Of File Systems and Storage Models

Chapter 4
Of File Systems and Storage

Models
Disks are always full. It is futile to try to get more disk space.
Data expands to fill any void. – Parkinson’s Law as applied to
disks
4.1 Introduction
This chapter deals primarily with how we store data. Virtually all computer
systems require some way to store data permanently; even so-called “diskless”
systems do require access to certain files in order to boot, run and be useful.
Albeit stored remotely (or in memory), these bits reside on some sort of
storage system.
Most frequently, data is stored on local hard disks, but over the last
few years more and more of our files have moved “into the cloud”, where
di↵erent providers o↵er easy access to large amounts of storage over the
network. We have more and more computers depending on access to remote
systems, shifting our traditional view of what constitutes a storage device.
74
CHAPTER 4. OF FILE SYSTEMS AND STORAGE MODELS 75
As system administrators, we are responsible for all kinds of devices: we

build systems running entirely without local storage just as we maintain the
massive enterprise storage arrays that enable decentralized data replication
and archival. We manage large numbers of computers with their own hard
drives, using a variety of technologies to maximize throughput before the
data even gets onto a network.
In order to be able to optimize our systems on this level, it is important for
us to understand the principal concepts of how data is stored, the di↵erent
storage models and disk interfaces. It is important to be aware of certain
physical properties of our storage media, and the impact they, as well as
certain historic limitations, have on how we utilize disks.
Available storage space is, despite rapidly falling prices for traditional
hard disk drives, a scarce resource1 . The quote a the beginning of this chap-
ter is rather apt: no matter how much disk space we make available to our
users, we will eventually run out and need to expand. In order to accom-
modate the ever-growing need for storage space, we use technologies such as
Logical Volume Management to combine multiple physical devices in a flex-
ible manner to present a single storage container to the operating system.
We use techniques such as RAID to increase capacity, resilience or perfor-
mance (pick two!), and separate data from one another within one storage
device using partitions. Finally, before we can actually use the disk devices
to install an operating system or any other software, we create a file system
on top of these partitions.
System administrators are expected to understand well all of these topics.
Obviously, each one can (and does) easily fill many books; in this chapter, we
will review the most important concepts underlying the di↵erent technologies
from the bottom up to the file system level. At each point, we will compare
and contrast traditional systems with recent developments, illustrating how
the principles, even if applied di↵erently, remain the same. For significantly
deeper discussions and many more details, please see the chapter references,
in particular the chapters on file systems in Silberschatz[8] and McKusick et
al.’s canonical paper on the Berkeley Fast File System[9].
1
We will take a closer look a the Total Cost of Ownership of a hard drive in Chapter
13. Suffice it to say that the purchase price of the storage device itself is only a fraction.
4.2 Storage Models

We distinguish di↵erent storage models by how the device in charge of keeping
the bits in place interacts with the higher layers: by where raw block device
access is made available, by where a file system is created to make available
the disk space as a useful unit, by which means and protocols the operating
system accesses the file system. Somewhat simplified, we identify as the main
three components the storage device itself, i.e. the actual medium; the file
system, providing access to the block level storage media to the operating
system; and finally the application software. The operating system managing
the file system and running the application software then acts as the agent
making actual I/O possible.
4.2.1 Direct Attached Storage

The by far most common way to access storage is so simple that we rarely
think about it as a storage model: hard drives are attached (commonly via
a host bus adapter and a few cables) directly to the server, the operating
system detects the block devices and maintains a file system on them and
thus allows for access with the smallest level of indirection. The vast majority
of hosts (laptops, desktop and server systems alike) all utilize this method.
The term used nowadays – Direct Attached Storage (DAS) – was e↵ectively
created only after other approaches become popular enough to require a
simple di↵erentiating name.
Figure 4.1a illustrates this model: all interfacing components are within
the control of a single server’s operating system (and frequently located
within the same physical case) and multiple servers each have their own
storage system. On the hardware level, the storage media may be attached
using a variety of technologies, and we have seen a number of confusing stan-
dards come and go over the years. The best choice here depends on many
factors, including the number of devices to be attached, driver support for the
connecting interface in the OS, and performance or reliability considerations.
We will touch upon all of these aspects throughout this chapter.
A server with a single hard drive (such as the one shown in Figure 4.1b is
perhaps the simplest application of this model. But it is not uncommon for
servers to have multiple direct attached devices, the configuration of which
then depends entirely on the next higher level of storage strategies. Individual
disks can simply be mounted in di↵erent locations of the file system hierarchy.
(a) Diagram (b) A Maxtor IDE Drive
Figure 4.1: Direct Attached Storage
Alternatively, multiple direct attached disks can be combined to create a

single logical storage unit through the use of a Logical Volume Manager
(LVM) or a Redundant Array of Independent Disks (RAID). This allows for
improved performance, increased amount of storage and/or redundancy. We
will discuss these concepts in more detail in Section 4.4.
Direct attached storage need not be physically located in the same case
(or even rack) as the server using it. That is, we di↵erentiate between in-
ternal storage (media attached inside the server with no immediate external
exposure) and external storage (media attached to a server’s interface ports,
such as Fibre Channel, USB etc.) with cables the lengths of which depend
on the technology used. External media allows us to have large amounts of
storage housed in a separate enclosure with its own power supply, possibly
located several feet away from the server. If a server using these disks suf-
fers a hardware failure, it becomes significantly easier to move the data to
another host: all you need to do is connect the cable to the new server.
Simple as this architecture is, it is also ubiquitous. The advantages of
DAS should be obvious: since there is no network or other additional layer
in between the operating system and the hardware, the possibility of failure
on that level is eliminated. Likewise, a performance penalty due to network
latency, for example, is impossible. As system administrators, we frequently
need to carefully eliminate possible causes of failures, so the fewer layers of
indirection we have between the operating system issuing I/O operations and
the bits actually ending up on a storage medium, the better.
At the same time, there are some disadvantages. Since the storage media
is, well, directly attached, it implies a certain isolation from other systems on
the network. This is both an advantage as well as a drawback: on the one
hand, each server requires certain data to be private or unique to its operating
system; on the other hand, data on one machine cannot immediately be made
available to other systems. This restriction is overcome with either one of the
two storage models we will review next: Network Attached Storage (NAS)
and Storage Area Networks (SANs).
DAS can easily become a shared resource by letting the operating system
make available a local storage device over the network. In fact, all network
file servers and appliances ultimately are managing direct attached storage
on behalf of their clients; DAS becomes a building block of NAS. Likewise,
physically separate storage enclosures can function as DAS if connected di-
rectly to a server or may be combined with others and connected to network
or storage fabric, that is: they become part of a SAN.2
4.2.2 Network Attached Storage

As the need for more and more data arises, we frequently want to be able
to access certain data from multiple servers. An old and still very common
example is to store all your users’ data on shared disks that are made available
to all clients over the network. When a user logs into hostA, she expects to
find all her files in place just as when she logs into hostB. To make this
magic happen, two things are required: (1) the host’s file system has to
know how to get the data from a central location and (2) the central storage
has to be accessible over the network. As you can tell, this introduces a
number of complex considerations, not the least of which are access control
and performance.
For the moment, let us put aside these concerns, however, and look at the
storage model from a purely architectural point of view: One host functions
as the “file server”, while multiple clients access the file system over the
network. The file server may be a general purpose Unix system or a special
network appliance – either way, it provides access to a number of disks or
2
Don’t worry, the various acronyms seem confusing as they are introduced, but hardly
anybody actually utters sentences containing more than one. One feels too silly doing so.
Figure 4.2: Three hosts using Network Attached Storage, or NAS
other storage media, which, within this system are e↵ectively direct attached
storage. In order for the clients to be able to use the server’s file system
remotely, they require support for (and have to be in agreement with) the
protocols used3 . However, the clients do not require access to the storage
media on the block level; in fact, they cannot gain such access.
From the clients’ perspective, the job of managing storage has become
simpler: I/O operations are performed on the file system much as they would
be on a local file system, with the complexity of how to shu✏e the data
over the network being handled in the protocol in question. This model
is illustrated in Figure 4.2, albeit in a somewhat simplified manner: even
though the file system is created on the file server, the clients still require
support for the network file systemthat brokers the transaction performed
locally with the file server.
3
The most common protocols in use with network attached storage solution are NFS
on the Unix side and SMB/CIFS on the Windows side. The Apple Filing Protocol (AFP)
is still in use in some predominantly Mac OS environments, but Apple’s adoption of Unix
for their Mac OS X operating system made NFS more widespread there as well.
In contrast to DAS, a dedicated file server generally contains significantly

more and larger disks; RAID or LVM may likewise be considered a require-
ment in this solution, so as to ensure both performance and failover. Given
the additional overhead of transferring data over the network, it comes as no
surprise that a certain performance penalty (mainly due to network speed or
congestion) is incurred. Careful tuning of the operating system and in par-
ticular the network stack, the TCPTCP window size, and the bu↵er cache
can help minimize this cost.
The benefits of using a central file server for data storage are immediate
and obvious: data is no longer restricted to a single physical or virtual host
and can be accessed (simultaneously) by multiple clients. By pooling larger
resources in a dedicated NAS device, more storage becomes available.
Other than the performance impact we mentioned above, the distinct
disadvantage lies in the fact that the data becomes unavailable if the net-
work connection su↵ers a disruption. In many environments, the network
connection can be considered sufficiently reliable and persistent to alleviate
this concern. However, such solutions are less suitable for mobile clients,
such as laptops or mobile devices, which frequently may disconnect from and
reconnect to di↵erent networks. Recent developments in the area of Cloud
Storage have provided a number of solutions (see Section 4.2.4), but it should
be noted that mitigation can also be found in certain older network file sys-
tems and protocols: the Andrew File System (AFS), for example, uses a local
caching mechanism that lets it cope with the temporary loss of connectivity
without blocking.
While network attached storage is most frequently used for large, shared
partitions or data resources, it is possible to boot and run a server entirely
without any direct attached storage. In this case, the entire file system,
operating system kernel and user data may reside on the network. We touch
on this special setup in future chapters.
4.2.3 Storage Area Networks

Network Attached Storage (NAS) allows multiple clients to access the same
file system over the network, but that means it requires all clients to use
specifically this file system. The NAS file server manages and handles the
creation of the file systems on the storage media and allows for shared access,
overcoming many limitations of direct attached storage. At the same time,
however, and especially as we scale up, our requirements with respect to
(a) Huawei Tecal RH 2288H V2 (b) NetApp FAS 3050c
Figure 4.3: NAS and SAN Hardware

NAS and SAN enterprise hardware. On the left, Huawei storage servers with
24 hard drives each and built-in hardware RAID controllers; on the right, a
NetApp Fabric Attached Storage device, also known as a “filer” with disk
enclosures commonly referred to as “shelves”.
storage size, data availability, data redundancy, and performance, it becomes

desirable to allow di↵erent clients to access large chunks of storage on a block
level. To accomplish this, we build high performance networks specifically
dedicated to the management of data storage: Storage Area Networks.
In these dedicated networks, central storage media is accessed using high
performance interfaces and protocols such as Fibre Channel or iSCSI, mak-
ing the exposed devices appear local on the clients. As you can tell, the
boundaries between these storage models are not rigid: a single storage de-
vice connected via Fibre Channel to a single host (i.e. an example of DAS) is
indistinguishable (to the client) from a dedicated storage device made avail-
able over a Storage Area Network (SAN). In fact, today’s network file servers
frequently manage storage made available to them over a SAN to export a
file system to the clients as NAS.
Figure 4.4 illustrates how the storage volumes managed within a SAN can
be accessed by one host as if it was direct attached storage while other parts
are made available via a file server as NAS to di↵erent clients. In order for the
Figure 4.4: A SAN providing access to three devices; one host accesses parts
of the available storage as if it was DAS, while a file server manages other
parts as NAS for two clients.
di↵erent consumers in a SAN to be able to independently address the distinct

storage units, each is identified by a unique Logical Unit Number (LUN). The
system administrator combines the individual disks via RAID, for example,
into separate volumes; assigning to each storage unit an independent LUN
allows for correct identification by the clients and prevents access of data
by unauthorized servers, an important security mechanism. Fibre Channel
switches used in SANs allow further partitioning of the fabric by LUNs and
subdivision into SAN Zones, allowing sets of clients specifically access to
“their” storage device only.
In this storage model the clients – the computers, file servers or other
devices directly attached to the SAN – are managing the volumes on a block
level (much like a physical disk, as discussed in Section 4.3.1). That is, they
need to create a logical structure on top of the block devices (as which the
SAN units appear), and they control all aspects of the I/O operations down
to the protocol. With this low-level access, clients can treat the storage like
any other device. In particular, they can boot o↵ SAN attached devices, they
can partition the volumes, create di↵erent file systems for di↵erent purposes
on them and export them via other protocols.
Storage area networks are frequently labeled an “enterprise solution” due
to their significant performance advantages and distributed nature. Espe-
cially when used in a switched fabric, additional resources can easily be
made available to all or a subset of clients. These networks utilize the Small
Computer System Interface (SCSI) protocol for communications between the
di↵erent devices; in order to build a network on top of this, an additional
protocol layer – the Fibre Channel Protocol (FCP) being the most common
one – is required. We will review the various protocols and interfaces in
Section 4.3.
SANs overcome their restriction to a local area network by further encap-
sulation of the protocol: Fibre Channel over Ethernet (FCoE) or iSCSI, for
example, allow connecting switched SAN components across a Wide Area
Network (or WAN). But the concept of network attached storage devices fa-
cilitating access to a larger storage area network becomes less accurate when
end users require access to their data from anywhere on the Internet. Cloud
storage solutions have been developed to address these needs. However, as
we take a closer look at these technologies, it is important to remember that
at the end of the day, somewhere a system administrator is in charge of mak-
ing available the actual physical storage devices underlying these solutions.
Much like a file server may provide NAS to its clients over a SAN, so do cloud
storage solutions provide access on “enterprise scale” (and at this size the use
of these words finally seems apt) based on the foundation of the technologies
we discussed up to here.
4.2.4 Cloud Storage

In the previous sections we have looked at storage models that ranged from
the very simple and very local to a more abstracted and distributed approach,
to a solution that allows access across even a Wide Area Network (WAN). At
each step, we have introduced additional layers of indirection with the added
benefit of being able to accommodate larger requirements: more clients, more
disk space, increased redundancy, etc.
We also have come full circle from direct attached storage providing block-
level access, to distributed file systems, and then back around to block-level
access over a dedicated storage network. But this restricts access to clients
on this specific network. As more and more (especially smaller or mid-sized)
companies are moving away from maintaining their own infrastructure to-
wards a model of Infrastructure as a Service (IaaS) and Cloud Computing,
the storage requirements change significantly, and we enter the area of Cloud
Storage4 .
The term “cloud storage” still has a number of conflicting or surprisingly
di↵erent meanings. On the one hand, we have commercial services o↵ering
file hosting or file storage services; common well-known providers currently
include Dropbox, Google Drive, Apple’s iCloud and Microsoft’s SkyDrive.
These services o↵er customers a way to not only store their files, but to
access them from di↵erent devices and locations: they e↵ectively provide
network attached storage over the largest of WANs, the Internet.
On the other hand we have companies in need of a more flexible storage
solutions than can be provided with the existing models. Especially the
increased use of virtualization technologies demands faster and more flexible
access to reliable, persistent yet relocatable storage devices. In order to meet
these requirements, storage units are rapidly allocated from large storage
area networks spanning entire data centers.
Since the di↵erent interpretations of the meaning of “cloud storage” yield
significantly di↵erent requirements, the implementations naturally vary, and
there are no current industry standards defining an architecture. As such,
we are forced to treat each product independently as a black box; system
administrators and architects may choose to use any number of combinations
of the previously discussed models to provide the storage foundation upon
which the final solution is built.
We define three distinct categories within this storage model: (1) services
that provide file system level access as in the case of file hosting services such
as those mentioned above; (2) services that provide access on the object level,
hiding file system implementation details from the client and providing for
easier abstraction into an API and commonly accessed via web services5 such
4
Large companies are of course also moving towards IaaS, only they frequently are
the ones simultaneously consuming as well as providing the service, either internally or to
outside customers.
5
“Web services” generally expose an API over HTTP or HTTPS using REpresentationsl
State Transfer (REST) or the Simple Object Access Protocol (SOAP).
Figure 4.5: A possible cloud storage model: an internal SAN is made available
over the Internet to multiple clients. In this example, the storage provider
e↵ectively functions as a NAS server, though it should generally be treated
as a black box.
as Amazon’s Simple Storage Service (S3), or Windows Azure’s Blob Storage;

and (3) services that o↵er clients access on the block level, allowing them to
create file systems and partitions as they see fit (examples include Amazon’s
Elastic Block Store (EBS) and OpenStack’s Cinder service).
All of these categories have one thing in common, however. In order
to provide the ability of accessing storage units in a programmatic way –
a fundamental requirement to enable the flexibility needed in demanding
environments – they rely on a clearly specified API. Multiple distributed
resources are combined to present a large storage pool, from which units are
allocated, de-allocated, re-allocated, relocated, and duplicated, all by way of
higher-level programs using well-defined interfaces to the lower-level storage
systems.
Customers of cloud storage solution providers reap a wealth of benefits,
including: their infrastructure is simplified through the elimination of storage
components; storage units are almost immediately made available as needed
and can grow or shrink according to immediate or predicted usage patterns;
applications and entire OS images can easily be deployed, imported or ex-

ported as virtual appliances.
Of course these benefits carry a cost. As usual, any time we add layers
of abstraction we also run the risk of increasing, possibly exponentially, the
number of ways in which a system can fail. Cloud storage is no exception:
by relying on abstracted storage containers from a third-party provider, we
remove the ability to troubleshoot a system end-to-end; by outsourcing data
storage, we invite a number of security concerns regarding data safety and
privacy; by accessing files over the Internet, we may increase latency and
decrease throughput; the cloud service provider may become a single point
of failure for our systems, one that is entirely outside our control.
4.2.5 Storage Model Considerations

As we have seen in the previous sections, the larger we grow our storage re-
quirements, the more complex the architecture grows. It is important to keep
this in mind: even though added layers of abstraction and indirection help
us scale our infrastructure, the added complexity has potentially exponential
costs. The more moving parts a system has, the more likely it is to break,
and the more spectacular its failure will be.
A single bad hard drive is easy to replace; rebuilding the storage array
underneath hundreds of clients much less so. The more clients we have, the
more important it is to build our storage solution for redundancy as well as
reliability and resilience.
System administrators need to understand all of the storage models we
discussed, as they are intertwined: DAS must eventually underly any storage
solution, since the bits do have to be stored somewhere after all; the concepts
of NAS permeate any infrastructure spanning more than just a few work
stations, and SANs and cloud storage combine DAS and NAS in di↵erent
ways to make storage available over complex networks.
At each layer, we introduce security risks, of which we need to be aware:
any time bits are transferred over a network, we need to consider the integrity
and privacy of the files: who has access, who should have access, how is the
access granted, how are clients authenticated, and so on. NAS and SAN
solutions tend to ignore many of these implications and work under the as-
sumption that the network, over which the devices are accessed, are “secure”;
access controls are implemented on a higher layer such as the implementa-
tion of the file system. Often times, access to the network in question implies
(a) Open Hard Disc Drive (b) SSD Mini PCIe

Card
Figure 4.6: An open PATA (or IDE) hard drive (left) and a Solid State Drive
(right). The HDD shows the rotating disk platters, the read-write head with
its motor, the disk controller and the recognizable connector socket.
access to the shared storage, even though layer-2 security mechanisms such
as IPsec may be combined with or integrated into the solution. Cloud stor-
age, on the other hand, has to directly address the problem of transmitting
data and providing access over untrusted networks and thus usually relies on
application layer protocols such as Transport Layer Security (TLS)/Secure
Sockets Layer (SSL).
We will touch on some of these aspects in future sections and chapters,
but you should keep them in mind as you evaluate di↵erent solutions for
di↵erent use cases. As we will see, in most cases the simpler model turns
out to be the more scalable and more secure one as well, so beware adding
unneeded layers of complexity!
4.3 Disk Devices and Interfaces

The di↵erent storage models we discussed in the previous sections are just a
means to access the storage devices in order to, well, store our data. Devices
or media used to store data include tape drives (for use with magnetic tape),
optical media such as CDs, various non-volatile memory based devices such as
flash drives, DRAM-based storage, and of course the hard-disk drive (HDD).
Even though Solid State Drive (SDD) o↵er significant advantages such as
lower power consumption and generally higher performance, the dominant

medium in use especially in enterprise scale storage solutions remains the
ubiquitous hard drive6 , storing data on rotating, magnetic platters (see Fig-
ure 4.6a). Understanding the physical structure of these traditional storage
devices is important for a system administrator, as the principles of espe-
cially the addressing modes and partition schemas used here come into play
when we look at how file systems manage data efficiently.
Hard drives can be made available to a server in a variety of ways. In-
dividual disks are connected directly to a Host Bus Adapter (HBA) using a
single data/control cable and a separate power cable. The traditional inter-
faces here are SCSI, PATA and SATA, as well as Fibre Channel.
SCSI, the Small Computer System Interface, has been around for over
25 years and has seen a large number of confusing implementations and
standards7 . Once the default method to connect any peripheral device using
long, wide and generally unwieldy ribbon cables, SCSI has now been largely
obsoleted by the Advanced Technology Attachment (ATA) standards. At
the same time, however, it lives on in the Internet Small Computer System
Interface (iSCSI), a standard specifying storage connections using the SCSI
command protocol over IP-based networks. iSCSI is a common choice in
storage area networks; as it uses TCP/IP over existing networks, it does
not require a dedicated storage network as is the case in traditional Fibre
Channel SANs.
The Parallel Advanced Technology Attachment (PATA) standard, also
frequently referred to as IDE (for Integrated Device Electronics, a reference
to the fact that the drive controller is included in the hard drive), uses a 40
or 80 wire ribbon cable and allows for a maximum number of two devices on
the connector (traditionally referred to as the master and slave; this is often
a source of confusion, as neither device takes control or precedence over the
other).
Faster throughput, smaller cables and support for hot-swapping, i.e. the
ability to replace a drive without having to shut down the operating system8 ,
6
Despite declining prices for SDDs, as of 2012 traditional hard drives remain notably
cheaper than SSDs and have higher storage capacity.
7
Di↵erent SCSI versions include such wonderful variations as Fast SCSI, Fast Wide
SCSI, Ultra SCSI, and Wide Ultra SCSI. None of which are to be confused with iSCSI, of
course.
8
Hot-swapping was a standard feature in many SCSI implementations. Many system
administrators in charge of the more powerful servers, using larger and more performant
were some of the advantages provided by the Serial ATA (SATA) interface.
A number of revisions and updates to the standard added more advanced
features and, most significantly, increasingly greater transfer speeds.
Most motherboards have integrated ATA host adapters, but a server can
be extended with additional HBAs via, for example, its PCI Express ex-
pansion slots; similarly, dedicated storage appliances make use of disk array
controllers to combine multiple drives into logical units (more on that in
Section 4.4). Fibre Channel HBAs finally allow a server to connect to a ded-
icated Fibre Channel SAN. All of these interfaces can be either internal (the
devices connected to the bus are housed within the same physical enclosure
as the server) or external (the devices are entirely separate of the server,
racked and powered independently and connected with suitable cables). In
the end, consider a host with a large amount of DAS and a NAS server man-
aging multiple terabytes of file space which is housed in a separate device and
which it accesses over a SAN: the main di↵erence lies not in the technologies
and protocols used, but in how they are combined.
Hands-on Hardware
I have found it useful to have actual hardware in class whenever
possible. In the past, I have brought with me di↵erent hard
drives by di↵erent manufacturers and with di↵erent interfaces
(PATA, SATA, SCSI); in particular, showing students the inside of a hard
drive, the rotating platters and the read-write arms has served as a great
illustration of the performance limiting factors, such as rotational latency,
seek time etc. Especially old and by now possibly obsolete hardware and
the contrast in which it stands to modern solutions is always investigated
with great interest.
Likewise, a large variety of cables, connectors and adapters help illustrate

the di↵erent standards and allow students to connect what they may
have seen or used themselves to what we talk about in class. (See also:
Problem 4)
SCSI drives when compared to the individual workstations at the time, were rather fond of
this: when a disk failed, it could be replaced without scheduling downtime for all services.
4.3.1 Physical Disk Structure

Let us open up a typical hard drive and take a look at the physical struc-
ture. This will help us better understand the concepts of partitions, how the
operating system calculates the disk geometry and allocates physical blocks,
how the file system manages cylinder groups, and why, for example, storing
data on the outer disk blocks could improve performance.
If you open a traditional hard-disk drive and manage not to destroy it
completely in the process, you will look at a number of magnetic platters on
a spindle together with the read-write heads. The other main components
include the disk motor (spinning the platters), an actuator (moving the read-
write heads) and the disk controller.
The platters are coated on both surfaces with magnetic particles, allowing
us to store on (and retrieve o↵) them data in the form or zeros and ones by
polarizing the relevant area. The read-write heads – one for each platter
surface – may be resting on the landing zone near the spindle; they do not
touch the platters when they are rotating. Instead, they hover just above the
platters and are moved radially across the disks by the actuator to position
them at the desired location. If the read-write heads are brought into contact
with the spinning platters, the results tend to be rather disastrous, as the
magnetic surface of the disk is damaged by this head crash.
The surface of each platter is divided into concentric rings, or tracks. Con-
gruent tracks on multiple platters form a three-dimensional cylinder. Within
a single cylinder, data can be read from all available surfaces without re-
quiring the read-write heads to move. This is why disk partitions comprise
multiple cylinder groups rather than, as we may sometimes imagine them to
be, pie wedges.
The tracks on each platter are in turn divided into a number of sectors.
If you were to divide a disk with concentric tracks by drawing straight lines
from the center to the outer edge, you would quickly realize that even though
you end up with the same number of such sectors on each track, the fields
created in this manner would be of varying sizes: the ones on the outer edge
would be significantly larger than the ones in the middle of the disc (see
Figures 4.7a and 4.7b).
As these fields represent the smallest addressable unit on a hard drive,
we would be wasting a lot of disk space. Instead, each sector is kept at a
fixed size9 and use a technique known as Zone Bit Recording to store more
9
The industry standard for hard drives used to be 512 bytes for many years; since
(a) Disk Structure: Cylinders/- (b) Physical Disk Structure with

Tracks, Heads, Sectors Zone Bit Recording
Figure 4.7: On the left: Illustration of tracks and sectors on a hard disk.
Note that for simplicity, sectors on the inside and outside of the platters are
of identical size. On disks with Zone Bit Recording (shown on the right),
this is no longer the case.
sectors on the outer tracks of each disc than on the inner area.
The total number of 512 byte sectors across all platters of the hard drive
thus define its total capacity. In order to access each storage unit, the read-
write head needs to be moved radially to the correct cylinder – a process we
call seeking – and the platter then spun until the correct sector is positioned
under it. In the worst case scenario, we just passed the sector we wish to
access and we have to perform a full rotation. Therefore, a drive’s perfor-
mance is largely defined by this rotational latency and the time to position
the read-write head (also known as seek time).
Since the motor of a drive may rotate the discs at a constant linear velocity
(versus constant angular velocity), the discs are in fact moving slower near
the spindle than on the outer part. This means that more sectors could
be read in a given time frame from the outside of the discs than from the
inside. In fact, it used to be not uncommon for system administrators to
partition their disks such that large, frequently accessed files would reside on
around 2010, a number of vendors have started creating hard drives with a physical block
size of 4096 bytes. Interestingly, file systems having standardized on 512 byte blocks tend
to divide these blocks and continue to present to the OS 512 byte “physical” blocks.
cylinders near the beginning (i.e. the outside) of the platters. Nowadays such
fine tuning may no longer be common, but instead people have started to
simply create a single partition occupying only about 25% of the disk at the
beginning and ignoring the rest. This technique, known as “short stroking”,
may seem wasteful, but the performance gain compared to the cheap prices
of today’s HDDs may make it actually worthwhile.
It is worth noting that the physical disk structure described here applies
only to traditional mechanical hard drives, not to Solid State Drives or other
storage media. Nevertheless, it is useful to understand the structure, as a
number of file system or partitioning conventions derive directly from these
physical restrictions.
True Hard Drive Capacity

Hard drive manufacturers use a di↵erent definition of what con-
stitutes a Gigabyte than is customary pretty much anywhere else:
normally, we define computer storage in terms of powers of two.
That is, a kilobyte is commonly understood to be 210 bytes, a
megabyte 220 and a gigabyte 230 . In contrast, the standard prefix
giga means 109 , and hard drive manufacturers use this meaning.
Some people di↵erentiate between these conflicting definitions by using
the term Gibibyte (GiB) for 230 bytes, but even though technically cor-
rect, this term has never really caught on.
As a result, a drive labeled as having a capacity of 500 GB will in fact

only be able to store around 465 GiB. The larger the drive, the bigger
the di↵erence - at 1TB (240 versus 1012 ), you are getting a full 10% less
storage than advertized!
4.4 Dividing and Combining Disks

Disk space is a finite and fixed resource. Each hard drive we purchase has a
given capacity, and it is up to the system administrators to use it efficiently.
In this section, we will take a look at the di↵erent methods to provide the
optimal amount of storage space to your systems. Sometimes this involves
dividing a single hard drive into multiple partitions; other times, we want
to combine multiple hard drives to increase capacity, yet retain a logical
view of the storage space as a single unit. Finally, the ways in which we
divide or combine disks have implications on system performance and data
redundancy.
4.4.1 Partitions
Now that we understand the physical layout of the hard disks, we can take
a look at how we partitions are created and used. As we noted in the pre-
vious section, a disk partition is a grouping of adjacent cylinders through
all platters of a hard drive10 . Despite this unifying principle, we encounter
a variety of partition types and an abundance of related terminology: there
are partition tables and disklables, primary and extended partitions; there
are whole-disk partitions, disks with multiple partitions, and some of the
partitions on a disk may even overlap.
Di↵erent file systems and anticipated uses of the data on a disk require
di↵erent kinds of partitions. First, in order for a disk to be bootable, we
require it to have a boot sector, a small region that contains the code that
the computer’s firmware (such as the Basic Input/Output System (BIOS))
can load into memory. In fact, this is precisely what a BIOS does: it runs
whatever code it finds in the first sector of the device so long as it matches
a very simple boot signature. That is, regardless of the total capacity of the
disk in question, the code that chooses how or what to boot needs to fit into
a single sector, 512 bytes. On most commodity servers11 , this code is known
as the Master Boot Record or MBR. In a classical MBR the last two bytes of
this sector contain the signature 0x55 0xAA; bootstrap code area itself takes
up 446 bytes, leaving 64 bytes of space. At 16 bytes per partition entry,
we can have at most four such BIOS partitions that the MBR can transfer
control to.
Sometimes you may want to divide the available disk space into more
than four partitions. In order to accomplish this, instead of four primary
10
Some operating systems such as Solaris, for example, have traditionally referred to
partitions as “slices”; some of the BSD systems also refer to disk “slices” within the context
of BIOS partitions. Unfortunately, this easily brings to mind the misleading image of a
slice of pie, a wedge, which misrepresents how partitions are actually laid out on the disk.
11
Even though many other hardware architectures used to be dominant in the server
market, nowadays the x86 instruction set (also known as “IBM PC-compatible” computers)
has replaced most other systems. For simplicity’s sake, we will assume this architecture
throughout this chapter.
# fdisk -l
Disk / dev / cciss / c0d0 : 73.3 GB , 73372631040 bytes
255 heads , 63 sectors / track , 8920 cylinders
Units = cylinders of 16065 * 512 = 8225280 bytes
Device Boot Start End Blocks Id System

/ dev / cciss / c0d0p1 * 1 2550 20482843+ 83 Linux
/ dev / cciss / c0d0p2 2551 7649 40957717+ 83 Linux
/ dev / cciss / c0d0p3 7650 7904 2048287+ 82 Linux swap
#
Listing 4.1: fdisk(8) sample invocation and output on a Linux system
partitions, the Master Boot Record (MBR) allows you to specify three pri-
mary and one so-called extended partition, which can be subdivided further
as needed. When the system boots, the BIOS will load the MBR code, which
searches its partition table for an “active” partition, from which it will then
load and execute the boot block. This allows the user to run multiple operat-
ing systems from the same physical hard drive, for example. BIOS partitions
are usually created or maintained using the fdisk(8) utility.
Disk and BIOS Limitations

In order to address sectors of a hard drive, computers used the
Cylinder-Head-Sector (CHS) scheme. Early BIOSes were only able
to address 1024 cylinders, 256 heads and 63 sectors/track. At the
same time, the ATA specification for IDE disks defined a limit of
65536 cylinders, 16 heads, and 256 sectors. As a result, the lowest
common denominator required addressing – and thus total disk
sizes – to be limited to:
1024(cylinders) ⇤ 16(heads) ⇤ 63(sectors) ⇤ 512(bytes/sector) = 504M B
Enhanced BIOSes removed some of these limitations, allowing for

disks with an at that time impressive size of approximately 8 GB
(1024 ⇤ 255 ⇤ 63 ⇤ 512). But they were only able to access the first 1024
cylinders of a disk. This meant that all the files required to boot the
OS (which then might be able to address more disk space) had to reside
within these cylinders. As a result, many Unix systems have a default
partitioning scheme including a small /boot partition created at the be-

ginning of the disk.
To overcome the limitations of CHS addressing, data blocks on the

disks were simply numbered and counted, yielding Logical Block Ad-
dressing (LBA). Using this method, IDE disks could thus address how-
ever many blocks it could store in the data type it used for this pur-
pose; the ATA specification used 28 bits, and thus imposed a limit of
(228 sectors ⇤ 512 bytes/sector)/10243 bytes/GB = 128 GB (or 137.4GiB
at 109 bytes/GiB).
In 2012, the current ATA/ATAPI specification uses 48 bits, limiting disk

size to 128 Petabytes (PB); however, some operating systems still use 32
bits to address sectors, so that on those systems the largest supported
disk size is 2 Terabyte (TB), yet another limitation we now run into and
which may seem laughably small in just a few years.
Just as the first sector of the disk contains the mbr, so does the first
sector of a BIOS partition contain a volume boot record, also known as a par-
tition boot sector. In this sector, the system administrator may have placed a
second-stage bootloader, a small program that allows the user some control
over the boot process by providing, for example, a selection of di↵erent ker-
nels or boot options to choose from. Only one partition is necessary to boot
the OS, but this partition needs to contain all the libraries and executables
to bootstrap the system. Any additional partitions are made available to the
OS at di↵erent points during the boot process.
In the BSD family of operating systems, the volume boot record contains
a disklabel, detailed information about the geometry of the disk and the par-
titions it is divided into. Listing 4.2 shows the output of the disklabel(8)
command on a NetBSD system. You can see the breakdown of the disk’s
geometry by cylinders, sectors and tracks and the partitioning of the disk
space by sector boundaries. This example shows a 40 GB12 disk containing
three partitions, a 10 GB root partition, a 512 MB swap partition and a data
partition comprising the remainder of the disk. Since the disk in question
12
(78140160 sectors ⇤ 512 bytes/sector)/(10243 bytes/GB) = 37.26 GB
Note the di↵erence in actual versus reported disk size:
(40 ⇤ 230 40 ⇤ 109 )/(10243 ) = 40 37.26 = 2.74
# dmesg | g r e p xbd3
xbd3 a t xenbus0 i d 3 : Xen V i r t u a l Block D e v i c e I n t e r f a c e
xbd3 : 24575 MB, 512 b y t e s / s e c t x 50331520 s e c t o r s
# d i s k l a b e l / dev / xbd3
t y p e : ESDI
d i s k : U n i v e r s a l Swap
label : disk1
b y t e s / s e c t o r : 512
s e c t o r s / t r a c k : 306432
tracks / cylinder : 1
s e c t o r s / c y l i n d e r : 306432
c y l i n d e r s : 255
t o t a l s e c t o r s : 78140160
rpm : 3600
interleave : 1
8 partitions :
# size offset f s t y p e [ f s i z e b s i z e cpg / s g s ]
a: 20972385 63 4 . 2BSD 4096 32768 1180 # ( Cyl . 0⇤ 20805)
b: 1048320 20972448 swap # ( Cyl . 20806 21845)
c: 78140097 63 unused 0 0 # ( Cyl . 0⇤ 77519)
d: 78140160 0 unused 0 0 # ( Cyl . 0 77519)
e: 56119392 22020768 4 . 2BSD 4096 32768 58528 # ( Cyl . 21846 77519)
#
Listing 4.2: disklabel(8) invocation and output on a NetBSD system
is actually a virtual disk, the information reported relating to the hardware,

such as the rpm rate, for example, is obviously wrong and should be ignored.
This serves as a reminder that a system administrator always may need to
have additional background knowledge (“this host is a virtual machine with
a virtual disk”) to fully understand the output of her tools in order to make
sense of it. Note also that some partitions overlap. This is not a mistake: the
BSD disklabel includes information about both the entire disk (partition ’d’)
as well as the BIOS partition assigned to NetBSD (partition ’c’, starting at
o↵set 63). Other partitions, i.e. partitions actually used by the OS, should
not overlap.
Being able to read and understand the detailed output of these commands
is important, and students are encouraged to practice making sense of dif-
ferent partition schemas across di↵erent operating systems (see exercise 5).
Dividing a single large disk into multiple smaller partitions is done for a
number of good reasons: if you wish to install multiple operating systems, for
example, you need to have dedicated disk space as well as a bootable primary
partition for each OS. You may also use partitions to ensure that data written
to one location (log files, for example, commonly stored under e.g. /var/log)
cannot cause you to run out of disk space in another (such as user data under
/home). Other reasons to create di↵erent partitions frequently involve the
choice of file system or mount options, which necessarily can be applied only
on a per-partition basis. We will discuss a number of examples in Section
4.5.
4.4.2 Logical Volumes

Using hard drives with fixed partitions may lead to a number of problems:
disk drives are prone to hardware failure, which may lead to data loss; reading
data from individual disk drives may su↵er performance penalties depending
on the location of the data on the disk; and, perhaps most notably, disks are
never large enough.
Partitions are a good way to divide a single disk and make available to
the OS smaller, distinct units of storage. But throughout the life time of our
systems, we frequently have a need to increase storage space – consider the
quote from the beginning of this chapter: “Data expands to fill any void.”
We can buy a larger hard drive and add it to the system and then migrate
data to it or simply mount it in a separate location13 , but neither solution
is particularly elegant. The problem here is that partitions are fixed in size
throughout their life time. Wouldn’t it be nice if we could simply combine
the storage from multiple disk drives and then create partitions spanning the
total?
Enter logical volumes. Much like a SAN may combine multiple storage
devices and make them available to its clients, so does a logical volume
manager (or LVM) combine multiple physical devices and present them to
the OS as a single resource. In this case, the abstraction occurs on the device-
driver level; some advanced file systems may include aspects of an LVM (see
ZFS’s zpool(1) command, for example).
The management of the locally attached storage devices via so-called log-
ical volume groups grants the system administrator a significant amount of
flexibility: the total storage space can easily be extended (and the file system,
if it supports this operation, grown!) by adding new disk drives to the pool;
file system performance can be improved by striping data across all available
drives; data redundancy and fault tolerance can be improved by mirroring
13
An approach aptly described as “JBOD”: “Just a Bunch Of Disks”.
Figure 4.8: Logical Volume Management lets you combine multiple physical
disks or partitions into a single volume group, from which logical volumes can
be allocated.
data on multiple devices. These last two advantages are also provided by the
RAID storage solution, which we will look at in more detail in the following
section.
LVMs manage and combine three distinct resources, as illustrated in Fig-

ure 4.8: physical volumes, logical volume groups and logical volumes. A phys-
ical volume can be a hard disk, a partition on a hard disk, or a logical unit
number of a local or remote device (such as SAN connected storage device).
The LVM divides the physical volumes into data blocks, so-called physical
extents, and allows the system administrator to group one or more of these
physical volumes into a logical volume group. In e↵ect, available storage
space is combined into a pool, where resources can dynamically be added or
removed. Out of such a volume group, individual logical volumes can then
be created, which in turn are divided into the equivalent of a hard disk’s
sectors, so-called logical extents. This step of dividing a logical volume group
into logical volumes is conceptually equivalent to the division of a single hard
drive into multiple partitions; in a way, you can think of a logical volume
as a virtual disk. To the operating system, the resulting device looks and
behaves just like any disk device: it can be partitioned, and new file systems
can be created on them just like on regular hard drive disks.
By creating logical extents as storage units on the logical volumes, the

LVM is able to grow or shrink them with ease (the data corresponding to
the logical extends can easily be copied to di↵erent physical extents and
remapped), as well as implement data mirroring (where a single logical extent
maps to multiple physical extents).
Since logical volume management is a software solution providing an ad-
ditional layer of abstraction between the storage device and the file system,
it can provide additional features both on the lower level, such as data mir-
roring or striping, as well as on a higher level, such as file system snapshots,
which an LVM can easily implement using a “copy-on-write” approach when
updating the logical extents. We will discuss this topic in more detail in our
chapter on backups in Section 13.1.
Note that some of the desirable features of logical volume management
are not entirely independent of the higher file system layer. For example, a
change in size of the logical volumes requires the file system to adapt, and
growing a file system to expand to consume additional disk space is often
easier than shrinking the file system. Hence, the choice to use an LVM needs
to be coupled with the “right” file system!
4.4.3 RAID
Logical Volume Managers provide a good way to consolidate multiple disks
into a single large storage resource from which individual volumes can be
created. An LVM may also provide a performance boost by striping data, or
redundancy by mirroring data across multiple drives.
Another popular storage technology used for these purposes is RAID,
which stands for Redundant Array of Independent Disks14 . Multiple disks
can be combined in a number of ways to accomplish one or more of these
goals: (1) increased total disk space, (2) increased performance, (3) increased
data redundancy.
Much like an LVM, RAID as well hides the complexity of the management
of these devices from the OS and simply presents a virtual disk comprised
of multiple physical devices. However, unlike with logical volume manage-
14
The acronym “RAID” is sometimes expanded as “Redundant Array of Inexpensive
Disks”; since disks have become less and less expensive over time, it has become more
customary to stress the “independent” part. A cynic might suggest that this change in
terminology was driven by manufacturers, who have an interest in not explicitly promising
a low price.
ment, a RAID configuration cannot be expanded or shrunk without data

loss. Furthermore, an LVMis a software solution; RAID can be implemented
on either the software or the hardware layer. In a hardware RAID solu-
tion, a dedicated disk array controller (such as a PCIcard) is installed in the
server, and interacted with via controlling firmware or host-level client tools.
As a software implementation, an LVM may provide RAID capabilities, as
may certain file systems. ZFS, originally developed at Sun Microsystems, for
example, includes subsystems that provide logical volume management and
o↵er RAID capabilities, as does Linux’s Btrfs.
When combining disks in a RAID configuration, the system administra-
tor has to carefully consider the advantages and disadvantages of the avail-
able options, so-called RAID levels, where performance, redundancy, and
efficiency of disk usage need to be weighed. The common RAID levels are
summarized in Table 4.1 (although many more combinations and obscure
levels exist); the by far most popular levels are: RAID 0, RAID 1 and RAID
5.
RAID 0 Striped array (block level), no parity or mirroring

RAID 1 Mirrored array, no parity or striping
RAID 2 Striped array (bit level) with dedicated parity
RAID 3 Striped array (byte level) with dedicated parity
RAID 4 Striped array (block level) with dedicated parity
RAID 5 Striped array (block level) with distributed parity
RAID 6 Striped array (block level) with double distributed parity
Table 4.1: A brief summary of standard RAID levels.
RAID 0
By writing data blocks in parallel across all available disks (see Figure 4.9a),
RAID 0 accomplishes a significant performance increase. At the same time,
available disk space is linearly increased (i.e. two 500 GB drives yield 1 TB
of disk space, minus overhead). However, RAID 0 does not provide any fault
tolerance: any disk failure in the array causes data loss. What’s more, as
you increase the number of drives, you also increase the probability of disk
failure.
RAID 1
This configuration provides increased fault tolerance and data redundancy
by writing all blocks to all disks in the array, as shown in Figure 4.9b. When
a disk drive fails, the array goes into degraded mode, with all I/O operations
continuing on the healthy disk. The failed drive can then be replaced (hot-
swapped), and the RAID controller rebuilds the original array, copying all
data from the healthy drive to the new drive, after which full mirroring
will again happen for all writes. This fault tolerance comes at the price of
available disk space: for an array with two drives with a 500 GB capacity,
the total available space remains 500 GB.
(a) Block-level striping (b) Mirroring
(c) Block-level striping with distributed parity
Figure 4.9: Three of the most common RAID levels illustrated. RAID 0
increases performance, as blocks are written in parallel across all available
disks. RAID 1 provides redundancy, as blocks are written identically to all
available disks. RAID 5 aims to provide increased disk space as well as
redundancy, as data is striped and parity information distributed across all
disks.
RAID 5
This level provides a bit of both RAID 0 and RAID 1: data is written
across all available disks, and for each such stripe the data parity is recorded.
Unlike in levels 2 through 4, this parity is not stored on a single, dedicated
parity drive, but instead distributed across all disks. See Figure 4.9c for an
illustration of the block distribution.
Since parity information is written in addition to the raw data, a RAID 5
cannot increase disk capacity as linearly as a RAID 0. However, any one of
the drives in this array can fail without impacting data availability. Again,
as in the case of a RAID 1 configuration, the array will go into degraded
mode and get rebuilt when the failed disk has been replaced. However, the
performance of the array is decreased while a failed drive remains in the
array, as missing data has to be calculated from the parity; the performance
is similarly reduced as the array is being rebuilt. Depending on the size of
the disks inquestion, this task can take hours; all the while the array remains
in degraded mode and another failed drive would lead to data loss.
Composite RAID and Failure Rates

Since a RAID configuration creates a (virtual) disk to present to the upper
layer (such as the OS), it should not come as a surprise that it is possible to
combine or nest RAID levels to achieve a combination of the benefits provided
by either one level; Table 4.2 lists some of the most popular combinations.
In each case, we increase fault tolerance at the expense of efficiency.
It is worth noting that the redundancy gained by combining RAID levels
is frequently overestimated. Here’s why: each drive has an expected failure
rate, each array a mean time to data loss. The “Independent” in RAID mis-
leads us to believe that the odds of any one disk failing is entirely unrelated
to the others. In reality, it is most common to use identical drives from the
same manufacturer if not even the same production batch in a given array.
That is, when a single drive fails and the array goes into degraded mode, the
RAID 0+1 Mirrored array of stripes

RAID 1+0 Striped array of mirrors
RAID 5+0 Striped array of multiple RAID 5
Table 4.2: Standard RAID levels can be combined.

probability of a second drive failing is actually higher than one would expect
with truly independent drives.
Secondly, the actions performed by the RAID on the disks needs to be
taken into consideration as well. When a disk in a RAID 5 array fails and
is replaced, all of the disks will undergo additional stress as all of the data
is read in order to rebuild the redundant array. This process may well take
several hours if not days – a very long period of time during which our data is
at increased risk. In other words, the failure of a single drive will necessarily
increase the probability of failure of the other drives15 .
When choosing a composite RAID architecture, it is well worth our time
to consider the Mean Time to Recovery (MTTR) in addition to the other
factors. From the moment that a disk failure is detected until the array has
been rebuilt, our data is at severe risk. In order to reduce the MTTR, many
system administrators deploy so-called hot spares, disks that are installed in
the server and known to the RAID controller, but that are inactive until a
disk failure is detected. At that time, the array is immediately rebuilt using
this stand-by drive; when the faulty disk has been replaced, the array is
already in non-degraded mode and the new disk becomes the hot spare.
Figure 4.10: An Apple Xserve RAID, a now discontinued storage device with
14 Ultra-ATA slots o↵ering Fibre Channel connectivity and implementing a
number of RAID levels in hardware as well as software across two independent
controllers.
15
RAID 6 improves this situation, at the expense of efficiency. Dropping prices have not
very surprisingly lead more and more enterprise environments to consider this trade-o↵
entirely acceptable.
Spectacular File System Confusion

In early 2006, while working at Stevens Institute of Technology
as a System Administrator, we bought a new Apple Xserve RAID
storage device (like the one seen in Figure 4.10) and populated its
14 disk slots with 400 GB drives, yielding (after RAID, file sys-
tem overhead and the previously discussed discrepancy in binary
versus decimal units) two RAID5 configurations with a capacity of 2.2
TB each, an impressive amount of a↵ordable storage space at the time.
The “left” RAID was dedicated to storing large amounts of video and
audio data made available to clients running Mac OS X. We connected
the RAID controller via Fibre Channel to a SAN switch, and from there
to an Apple Xserve network server, which managed the HFS+ file system
on this storage component.
The second 2.2 TB of storage space, the ”right” side of the array, was
meant to become the central data space for all workstations in the Com-
puter Science and Mathematics departments as well as their laboratories.
Up until then, this file space had been provided via NFS from a two-mod-
ule SGI Origin 200 server running IRIX, managing a few internal SCSI
disks as well as some Fibre Channel direct attached storage. We intended
to migrate the data onto the XServe RAID, and to have it served via a
Solaris 10 server, allowing us to take advantage of several advanced fea-
tures in the fairly new ZFS and to retire the aging IRIX box.
Neatly racked, I connected the second RAID controller and the new So-
laris server to the SAN switch, and then proceeded to create a new ZFS
file system. I connected the Fibre Channel storage from the IRIX server
and started to copy the data onto the new ZFS file system. As I was
sitting in the server room, I was able to see the XServe RAID; I noticed
the lights on the left side of the array indicate significant disk activity,
but I initially dismissed this as not out of the ordinary. But a few seconds
later, when the right side still did not show any I/O, it dawned on me:
the Solaris host was writing data over the live file system instead of onto
the new disks!
I immediately stopped the data transfer and even physically dis-

connected the Solaris server, but the damage was done: I had
inadvertently created a new ZFS file system on the disks already
containing (and using) an HFS+ file system! As it turns out, I
had not placed the Solaris server into the correct SAN zone on the
Fibre Channel switch, meaning the only storage device it could
see was the left side of the array. But since both sides of the array were
identical (in size, RAID type, manufacturer), it was easy for me not to
notice and proceed thinking that due to proper SAN zoning, it was safe
for me to write to the device.
Now it was interesting to note that at the same time as I was overwriting
the live file system, data was still being written to and read from the
HFS+ file system on the Apple server. I was only able to observe inter-
mittent I/O errors. Thinking I could still save the data, I made my next
big mistake: I shut down the Apple server, hoping a clean boot and file
system check could correct what I still thought was a minor problem.
Unfortunately, however, when the server came back up, it was unable to
find a file system on the attached RAID array! It simply could not iden-
tify the device. In retrospect, this is no surprise: the Solaris server had
constructed a new (and di↵erent) file system on the device and destroyed
all the HFS+ specific file system meta data stored at the beginning of
the disks. That is, even though the blocks containing the data were likely
not over written, there was no way to identify them. After many hours
of trying to recreate the HFS+ meta data, I had to face the fact this
was simply impossible. What was worse, I had neglected to verify that
backups for the server were done before putting it into production use –
fatal mistake number three! The data was irrevocably lost; the only plus
side was that I had learned a lot about data recovery, SAN zoning, ZFS,
HFS+ and file systems in general.
4.5 File Systems

When the operating system is installed, it will create a file system on the
specified disk(s). As we have seen, these disks may be made available to the
OS in a variety of ways, but in the end, the OS does not distinguish between
a DAS device such as a hard drive disk, a volume created by an LVM, a disk
represented by a RAID or one backed by a storage device over a SAN. This
method of hiding hardware implementation choices or details allows the OS
to choose amongst a number of di↵erent file systems, each with their own
unique strengths, disadvantages and performance impacting tuning options.
Reduced to its most basic tasks, a file system is responsible for storing,
managing and updating data on the storage device in question. To this end,
it needs to manage the available space and be able to store files as well as their
attributes (i.e. metadata) in a hierarchy that allows both the user and the OS
to efficiently retrieve them, as well as ensure some level of file and file system
integrity. In addition, since Unix is a multi-user OS, the file system depends
on means to identify and distinguish between di↵erent users; access control
and the concepts of file permissions are a basic requirement. What’s more,
Unix systems frequently represent as “files” many things that us humans
may not usually think of as such16 . This leads us to special purpose file
systems that really only provide a file I/O API to their respective resources:
the procfs file system, representing information about the system’s processes
and the devfs file system, a virtual file system acting as an interface to device
drivers are two examples.
4.5.1 File System Types

In this section, we will take a look at some of the file systems commonly in use
on Unix systems. For the most part, these are disk file systems, meaning that
they are designed to manage hard disk storage. As we’ve seen in previous
sections, the physical layout of a hard drive influences partitioning schemas,
but the choice and creation of the operating system’s file system may also
depend on the underlying hardware.
16
The “Plan 9 from Bell Labs” research operating system took the initial Unix mantra
of “everything is a file” much further: all objects are either files or file systems, commu-
nicating with each other via the 9P protocol. Network connections, communications with
the kernel’s drivers, interprocess communication etc. all are performed via standard I/O
calls.
As we discuss these file systems, we will find that the approach to store
both the actual file data as well as the metadata, the information associated
with the files, di↵ers significantly. We will also notice – as a recurring pattern
throughout this book – how di↵erent layers of abstraction help us improve
portability and provide a consistent interface, as standard file I/O semantics
can remain the same across di↵erent file systems.
Disk File Systems

A Disk File System is probably what most people – at least those who would
think about these kinds of things to begin with – have in mind when they
think about what a file system does. It manages block device storage, stores
“files” in “directories”, maintains a file system hierarchy, controls file meta-
data, and allows for simple I/O via a few basic system calls. The canonical
file system for the Unix family of operating systems is, of course, the Unix
File System (UFS), also known as the Berkeley Fast File System (FFS). We
will look at the UFS in detail in Section 4.7, but for now suffice it to say that
this file system implementation and its use of boot blocks, superblocks, cylin-
der groups, inodes and data blocks is repeated and reflected in many other
Unix file systems, including for example ext2, for a long time the default file
system for Linux systems.
These traditional file systems su↵er a notable drawback in the case of
unexpected power failure or a system crash: some of the data may not have
been committed to the disk yet, causing a file system inconsistency. When
the OS boots up again, it will need to perform time consuming checks (see
fsck(8)) and any problems found may not actually be recoverable, leading to
possible data loss. In order to address these problems, a so-called journaled
file system might choose to first write the changes it is about to make to
a specific location (the journal) before applying them. In the event of a
crash, the system can then simply replay the journal, yielding a consistent
file system in a fraction of the time it would take a traditional file system to
traverse the entire hierarchy to ensure consistency.
There are many di↵erent journaled file systems, including a number of
commercial implementations (such as Apple’s HFS+, IBM’s JFS, or SGI’s
XFS) and open source variants (including ext3, ext4 and reiserfs for Linux
and updates to UFS/FFS providing support for journaling or logging).
Despite the di↵erences in how data is ultimately written to the disk, the
fundamental concepts of how data and metadata are managed, the use of
the inode structures, or the Virtual File System layer to allow the kernel to
support multiple di↵erent file systems remains largely the same across the
di↵erent Unix operating and file systems.
Distributed File Systems

Unlike a disk file system, which provides access to the data it holds only to the
processes running in the operating system the disk is attached to, a distributed
file system may allow di↵erent client systems to access a centralized storage
resource simultaneously, thus forming the basis for any NAS solutions. As we
discussed in Section 4.2.2, this model implies that clients no longer operate
on the block level of the storage – this access has been abstracted on a higher
level and the client OS thus requires in-kernel support for the file system in
question.
Sun Microsystems’ Network File System (NFS), created in 1985, was the
first such file system utilizing the Remote Procedure Calls (RPC) via the
Internet Protocol (IP)Internet Protocol for the server to communicate with
the clients, and remains to this day the standard distributed file system in the
Unix world. Other notable variations include the Andrew File System (AFS)
and its descendent, Coda, developed at Carnegie Mellon University, tightly
integrated with the Kerberos authentication protocol, and backed by a local
cache, making it particularly tolerant of network interruptions, as well as
the Server Message Block/Common Internet File System (SMB/CIFS), the
predominant network file system in the Windows world, made accessible to
Unix systems via the “Samba” implementation of these protocols.
Within the area of massive scale Internet systems, a number of specialized,
distributed file systems have emerged, most notably the Google File System[3]
and the Hadoop Distributed File System[4] (the latter being an open source
implementation modeled after the former). These file systems provide highly
scalable fault tolerance while at the same time providing high performance
on low-cost commodity hardware.
As usual, di↵erent problems call for di↵erent solutions, and despite their
significant capabilities, these file systems may not necessarily be suitable
for your requirements. In fact, making use of these capabilities requires a
significant investment – both in the time and expertise required to manage
them as well as in the hardware required to really get anything out of them.
Other distributed file systems may have a di↵erent focus, leading to in-
teresting and entirely distinct properties: the Tahoe Least-Authority File
System[5], for example, provides data confidentiality through the use of cryp-
tography, thus allowing complete decentralization over the (public) Internet.
On the other hand, the advent of cloud computing has further blurred the
boundaries of where a distributed file system ends and where a network
service begins: Amazon’s Simple Storage Service (S3), for example, despite
being an API driven web service, might as well be considered a distributed
file system, but then so could online storage providers, such as Dropbox, who
build their solutions on top of Amazon’s service. It will be interesting to see
how these distributed file systems or data stores evolve with time.
“Other” File Systems

In the Unix world, there exists an old mantra: “Everything is a file.” That is,
the simple API defining file I/O has proven so useful that a number of non-file
“things” have come to implement this interface as well: not only can regu-
lar files be accessed using the open(2), read(2), write(2) and close(2)
system calls17 , but so can network sockets and other interprocess communi-
cation endpoints, character and block devices, so-called pseudo-devices (such
as /dev/null), and various other “special files”.
Similarly, a number of pseudo- and virtual file systems have been created
to use the same API and make resources available using file-like structures.
The access points for these structures becomes a mount point in the file
system hierarchy, et voilà, these resources can be accessed, processed and
manipulated using common Unix tools. Examples include the devfs virtual
file system (used to access special and general devices), the procfs pseudo
file system (providing information about system information and individual
processes), the UnionFS (a method of layering two regular file systems on
top of each other, providing the union of the two hierarchies to the user),
and tmpfs (a method of creating a file system in virtual memory, granting
significant performance benefits).
In addition to these cases, there are also a number of file systems that
have been created to meet very specific requirements, or that e↵ectively func-
tion as an overlay on top of an application or protocol in order to facilitate
local file access to remote resources, such as, for example, the SSH file sys-
tem. But file systems are usually implemented in the operating system’s
17
All of these system calls operate on a small non-negative integer, the so-called file
descriptor or file handle. The old expression should therefore more accurately be cited as
“Everything is a file descriptor.”
kernel space: file I/O system calls do happen here and devices and certain
other resources can only be controlled by the kernel; likewise, mounting or
unmounting file systems requires superuser privileges. But not all file sys-
tems are actually accessing any of the resources protected by the kernel. In
order to facilitate non-privileged access to file system interfaces, some Unix
systems have created kernel modules that provide a method to implement
virtual “File Systems in Userspace” (known as FUSE).
All of these file systems that aren’t actually file systems but rather an ab-
straction of other resources into a file API illustrate how much the concept of
simplicity permeates the Unix culture. At this point, few system administra-
tors would wish to part with the convenience of, for example, procfs – in fact,
its availability is almost universally assumed. For a polyglot administrator,
however, it is important to be familiar with the di↵erent pseudo- and virtual
file systems available on the di↵erent Unix versions and, more importantly,
know how to access the resources they represent in their absence.
4.6 File System Layout

The traditional Unix file system layout is organized as a tree-like hierarchy,
rooted at / (“slash”). Under this directory reside a number of standard top-
level subdirectories, each with a well-defined purpose and containing further
subdirectories, forming a hierarchy described in the hier(7) manual page.
Unix systems can make use of multiple di↵erent file systems on multiple
partitions at the same time. Each may be attached or “mounted” under any
directory (or “mount point”) of the file system hierarchy with the exception
of the root file system: the file system containing the basic operating system
and from which the system was booted, which must be mounted at /.
Figure 4.10 illustrates a subtree of the typical Unix hierarchy with di↵er-
ent components residing on separate disks. Each node in the tree represents
a directory entry used to map file names to inodes; nodes with children rep-
resent directories, leaf nodes represent regular files. Traversing the tree from
the root to any given node constructs a so-called “pathname” consisting of
directory names concatenated with a path separator, “/”, followed by the
file name.
File names may contain any character except for this separator or the
NUL character (\0), used in the the C programming language to terminate
Figure 4.10: The Unix file system is a tree-like structure, rooted at /; di↵erent
file systems can be attached at di↵erent directories or mount points. In this
illustration, /home and /usr reside on separate disks from /.
a string.18 Most Unix systems impose a maximum file name length of 255
bytes and a maximum pathname length of 1024 bytes; however, these are file
system and OS specific limits.
Every Unix process has the concept of a current working directory – run
the pwd(1) command or type “echo $PWD” in your shell to display where in
the file system hierarchy you currently are. Pathnames may either be relative
to this location, or, if they begin with a /, absolute.
Each directory contains at least two entries: “.” (dot), a name for the
current working directory and “..” (dot-dot), a name for the parent direc-
18
It is a common mistake to assume that a file name contains only printable characters.
The Unix system does not impose many restrictions on how a user might choose to name
a file, and file names containing e.g. control characters such as a carriage return (“
n”), while confusing on a line-bu↵ered terminal, are possible.
$ pwd
/home/ jschauma # pwd ( 1 ) w r i t e s t h e a b s o l u t e pathname
$ echo h e l l o > f i l e # c r e a t e ” f i l e ” in the current d i r e c t o r y
$ cd / u s r / s h a r e / doc # cd ( 1 ) u s i n g an a b s o l u t e pathname
$ pwd
/ u s r / s h a r e / doc # no s u r p r i s e h e r e
$ c a t /home/ jschauma / f i l e # now u s i n g an a b s o l u t e pathname
hello
$ cd . . / . . / . . / home/ jschauma # cd ( 1 ) u s i n g a r e l a t i v e pathname
$ pwd
/home/ jschauma
$ cat f i l e # ” . / f i l e ” would a l s o work
hello
$
Listing 4.3: Absolute pathnames begin with a / and are resolved from the
root of the file system; relative pathnames are resolved from the current
working directory.
tory.19 Since relative pathnames are resolved from within the current working
directory, the same file can be referred to by di↵erent names, as shown in the
examples in Listing 4.3.
4.7 The Unix File System

As we discuss the general concepts of file systems in the Unix world, it is
impossible to avoid a much closer look at the Unix File System (UFS), the
initial standard file system for a number of both commercial and open source
Unix versions for decades. UFS, also known as the Berkeley Fast File System
or FFS, created by Marshall Kirk McKusick[9] and others at the University
of California in Berkeley in 1984, was a reimplementation of the original file
system provided by Bell Lab’s UNIX V7.
These improvements and changes introduced most notably the concept of
cylinder groups, providing a more equal distribution of the file metadata on
the physical disk, thus minimizing seek time. The Unix File System therefore
consisted of the following major components, as illustrated in Figure 4.11:
• A superblock. This block contains all the system’s crucial informa-

tion, such as the number and location of cylinder groups used as well as
19
As a special case, within the root directory both “.” and “..” refer to the same
directory, i.e. /.
Figure 4.11: A disk may be divided into multiple partitions; a partition

may contain a file system with multiple cylinder groups; each cylinder group
contains some file system meta data as well as inode and data blocks.
the location of the inode and data blocks. As this information is critical
for the operation of the file system and corruption of this block would
be disastrous, it is replicated and stored in a number of (predictable)
locations. This allows the super user to repair a corrupted file system
by pointing to an alternate superblock.
• A number of cylinder groups, which break the large file system into
more manageable chunks by distributing meta data evenly across the
physical partition.
• A number of inode maps (one for every cylinder group), pointing to

the cylinder group’s inode blocks, which in turn contain the metadata
associated with the files.
• A number of block bitmaps (one for every cylinder group), pointing

to the cylinder group’s data blocks, which in turn contain the actual
file data.
The distribution of this data across multiple cylinder groups illustrates

the tight relationship that the file system had with the physical properties
of a hard disk drive; this layout remains in place nowadays, even though on
solid state drives, for example, we no longer su↵er performance penalties due
to seek time.
It is important to note that the data blocks used by the file system are
di↵erent from the physical blocks of the hard disk. The latter are, as we
discussed in Section 4.3.1, 512 bytes in size (or, in more recent drives, 4096
bytes); the former – called the logical block size – can be decided on by the
system administrator at file system creation time. UFS uses a minimum log-
ical block size of 4096 bytes, and defaults to larger block sizes based on the
overall size of the file system. Likewise, the number of inodes in a file system
is fixed once the file system has been created. Here, UFS defaults to an inode
density of one inode per 2048 bytes of space for small file systems – consult
the newfs(8) manual page for this and other parameters to define and tune
the file system at its creation time.
Let us dive a little bit deeper and think about how the Unix File System
manages disk space: A file system’s primary task is to store data on behalf
of the users. In order to read or write this data, it needs to know in which
logical blocks it is located. That is, the file system needs a map of the blocks,
a way to identify and address each location. This is accomplished by way
of the inode and data block maps: the total number of inodes represents the
total number of files that can be referenced on this file system, while the data
blocks represent the space in which the file data is stored.
As we noted, a data block is, necessarily, of a fixed size. That means
that if we wish to store a file that is larger than a single block, we have to
allocate multiple blocks and make a note of which blocks belong to the given
file. This information is stored as pointers to the disk blocks within the inode
data structure.
Unfortunately, however, not all files will be multiples of the logical block
size. Likewise, it is possible that files will be smaller than a single block.
In other words, we will always end up with blocks that are only partially
allocated, a waste of disk space. In order to allow more efficient management
of small files, UFS allowed a logical block to be divided further into so-called
fragments, providing for a way to let the file system address smaller units of
storage. The smallest possible fragment size then is the physical block size
of the disk, and logical blocks are only fragmented when needed.
A pointer to the data blocks and fragments allocated to a given file is
stored in the inode data structure, which comprises all additional information
about the file. This allows for an elegant separation of a file’s metadata,
$ ls ai /
2 . 5740416 cdrom 5816448 libexec 2775168 stand
2 .. 1558656 dev 1862784 mnt 2166912 tmp
3003280 . c s h r c 988439 emul 4 netbsd 3573504 usr
3003284 . p r o f i l e 342144 etc 1824768 proc 3497472 var
3421440 a l t r o o t 1026432 home 798336 rescue
5702400 b i n 3763584 lib 3003264 root
3 boot . c f g 2204928 libdata 5588352 sbin
$
Listing 4.4: Use of the ls(1) command on a NetBSD system to illustrate how
file names are mapped to inode numbers in a directory. (Note that in the
root directory both ’.’ and ’..’ have the same inode number as in this special
case they actually are the same directory.)
which takes up only a fixed and small amount of disk space, and its contents.
Accessing the metadata is therefore independent of the file size (itself a piece
of metadata), allowing for efficient and fast retrieval of the file’s properties
without requiring access of the disk blocks. Other pieces of information
stored in the inode data structure include the file’s permissions, the numeric
user-id of the owner, the numeric group-id of the owner, the file’s last access,
modification and file status change times20 , the number of blocks allocated
for the file, the block size of the file system and the device the file resides on,
and of course the inode number identifying the data structure.
Now humans tend to be rather bad at remembering large numbers and
prefer the use of strings to represent a file, but the one piece of information
that is not stored in the inode is the file name. Instead, the Unix File System
allows for a mapping between a file name and its unique identifier – its inode
– through the use of directories. A directory really is nothing but a special
type of file: it has the same properties as any other file, but the data it
holds is well-structured (in contrast to the byte stream contained in so-called
“regular” files) and consists of inode number and file name pairs. You can
use the ls(1) command to illustrate this – see Listing 4.4.
A mapping of an inode number to a file name is called a hardlink, even
though humans tend to prefer the term “file name”. An inode can be accessed
by more than just one file name: within a single file system, each file may
20
A file’s ctime, it’s time of last file status change, is frequently misinterpreted to be the
file’s creation time; most file systems, including UFS, do not store this kind of information.
Instead, the ctime reflects the last time the meta information of a file was changed.
be referenced by multiple names in di↵erent directories, and the number of

hardlinks for a given file is stored in the inode data structure as well. But
since each file is identified by its inode number, and since multiple file systems
can be mounted in a running operating system, it is possible that two files
on two di↵erent file systems have the same inode number. As a result, a file
is really only uniquely identified by its combination of file system device and
inode number. What’s more, at times it may be useful to create a pointer to
a file residing on a di↵erent file system, but as just explained, creating a hard
link across file system boundaries is impossible. To overcome this limitation,
so-called “symbolic links” were invented: a symbolic link (often abbreviated
as “symlink”) is a special kind of file that contains as its only contents the
path name of the file it references. Almost all operations on a symlink will
then be redirected to the target file it points to.
In addition to regular files (hard links), directories and symbolic links,
the following file types are supported by the traditional Unix file system:
• Block special devices – an interface for disk-like devices, providing bu↵ered

and non-sequential I/O.
• Character special devices – an interface for communication devices such

as keyboards, mice, modems, or terminals, providing unbu↵ered I/O. A
number of virtual, so-called pseudo-devices such as /dev/null, /dev/zero,
or /dev/urandom, for example, are also accessed as character special
devices.
• Named pipes or FIFOs – another inter-process communications end-

point in the file system. This type of file represents a manifestation
of the traditional Unix pipe in the file system name space; I/O is per-
formed using the same methods as for any regular file. As data is
merely passed between processes, a FIFO always has a file size of zero
bytes.
• Unix domain sockets – an inter-process communications endpoint in the

file system allowing multiple processes with no shared ancestor process
to exchange information. Communication happens via the same API
as is used for network sockets.
The file type as well as its permissions and all the other properties of
a file can be inspected using the stat(1) command (see Listing 4.5 for an
$ stat / etc / passwd

File : ‘/ etc / passwd ’
Size : 2062 Blocks : 8 IO Block : 4096
regular file
Device : ca02h /51714 d Inode : 725181 Links : 1
Access : (0644/ - rw -r - -r - -) Uid : (0/ root ) Gid : (0/ root )
Access : 2011 -08 -26 23: 45 :3 1. 00 00 00 00 0 -0400
Modify : 2011 -08 -26 23: 45 :3 1. 00 00 00 00 0 -0400
Change : 2011 -08 -26 23: 45 :3 1. 00 00 00 00 0 -0400
$
Listing 4.5: Sample output of the stat(1) command on a Linux system show-
ing the various pieces of information stored in the inode data structure for
the file “/etc/passwd”.
example), or, perhaps more commonly, using the ls(1) command (as illus-
trated in Figure 4.12). This command is so frequently used and its output
so ubiquitous that any system administrator can recite the meaning of the
fields in their sleep. The semantics and order in which permissions are ap-
plied, however, include a few non-obvious caveats, which is why we will look
at the Unix permissions model in more detail in Chapter 6.
Over the last 30 years, UFS has served as the canonical file system for
almost all Unix versions. With time, a number of changes have become
necessary: more fine-grained access controls than the traditional Unix per-
missions model allows were made possible via file system extensions such as
ACLs; larger storage devices required not only updates to the block address-
ing schemas, but also to the di↵erent data types representing various file
system aspects; today’s huge amounts of available data storage have made
log-based file systems or journaling capabilities a necessity, and massively dis-
tributed data stores pose entirely di↵erent requirements on the underlying
implementation of how the space is managed.
Yet through all this, as enhancements have been made both by commer-
cial vendors as well as by various open source projects, the principles, the very
fundamentals of the Unix File System have remained the same: the general
concept of the inode data structure and the separation of file metadata from
the data blocks have proven reliable and elegant in their simplicity. At the
same time, this simplicity has proven to yield scalability and adaptability:
Figure 4.12: The default output of the ls -l command includes most of the
metadata of a given file.
the persistent idea that “everything is a file”, and that files simply store bytes
and thus have no inherent structure (unlike a database or certain archives)
have allowed a surprising flexibility and lead to the creation of many pseudo-
or virtual file systems providing a simple API and User Interface (UI) to any
number of resources.
4.8 Conclusions
Throughout this chapter, we have built our understanding of file systems
and storage models from the ground up. We have seen how simple concepts
are combined to construct increasingly complex systems – a pattern that
weaves like a red thread through all areas we cover. We noted the circular
nature of technology development: the simple DAS model repeats itself,
albeit more complex and with additional layers, in common SANs, much as
network attached storage utilizes both DAS and SAN solutions, yet is taken
to another extreme in cloud storage solutions.
Being aware of the physical disk structure helps us understand a file sys-
tem’s structure: we realize, for example, that concentric cylinder groups make
up partitions, and that the location of data blocks within these cylinders may
have an impact on I/O performance. What’s more, the concepts of file sys-
tem blocks become clearer the further we deepen our knowledge of both the
logical and physical components, and the distinction of metadata from the
actual contents of a file allow us to explain how the various file system related
Unix tools operate on a fairly low level as well as how to tune these values
at file system creation time. As system administrators, this understanding
is crucial.
But what about our earlier point of noting the three pillars of strong sys-
tem architecture and design: Scalability, Security, Simplicity. What role do
these play in the context of file systems and storage models?
Scalability is defined by our aptitude to meet changing requirements. As

far as file systems are concerned, we need to be able to provide sufficient
storage that meets our performance and rapidly increasing space require-
ments. One might be tempted to jump to a quick conclusion and choose the
superficially “most flexible” solution, which nowadays often appears to be
a cloud based storage model. However, it behooves us to take a step back
and consider the other two major design decisions, simplicity and security,
which help us cut through complexities and clarify just how flexible such a
model might be. Reliance on a third party provider for data storage, for
example, incurs a number of hidden costs: bandwidth requirements increase,
data redundancy and duplication need to be weighed, backup and data mi-
gration strategies change radically, and vendor lock-in may in fact eliminate
true flexibility.21 .
So what makes for a scalable storage solution? On the one hand, we need
to be able to add disk space on demand, so we either require a file system that
can grow as needed, or an efficient way to synchronize multiple file systems.
This already illustrates a trade-o↵ in complexity and the solutions depend in
part on our network architecture: within a single system, the use of an LVM
or a file system that allows the pooling of multiple resources seems a good
choice, but how do you make that available to multiple clients? If you have
hosts in distinct geographical locations, connected only via the Internet (an
inherently untrustworthy network), how do you allow for shared access?
The more you think about these requirements, the more you will find
yourself weighing conflicting approaches. You will go back and forth be-
tween the simplest possible solution (for example, all hosts write data only
to direct attached storage) and the more complex (shared storage available
21
Depending on the size of your organization, these costs may in fact not be hidden at
all: at some point, you may find yourself providing “cloud storage” to your organization,
and all of a sudden you are the owner of that infrastructure.
via a distributed file system or a cloud architecture). But it becomes increas-

ingly hard to define “simple”, as each solution requires considerations outside
of the initial problem domain: hosts writing to DAS is efficient, but data
synchronization across all hosts becomes increasingly complicated; a shared
storage pool available over the network solves that problem, but incurs a
high cost of complexity and dependency on the supporting infrastructure.
What’s more, the security aspects of the file system and the features it
provides are manifold: as Unix was designed from the beginning as a multi-
user operating system, the capabilities to distinguish between di↵erent users
were built into the system and is reflected in the file permissions. But file
system security goes beyond this basic level, especially when we consider
access over the network. Every distributed file system has to solve the prob-
lem of deciding whether or not a remote system should be granted access
to a given resource, and how such a system should authenticate. The ac-
cess models range from simplistic (if you are on the local network, you are
granted access) to the more sophisticated (a client has to present some sort
of time-limited authentication token to gain access).
But a system administrator’s professional paranoia dictates digging even
deeper. In the case of distributed file systems, of storage area networks or of
remotely hosted, cloud-based storage solutions, we not only have to consider
how a remote client authenticates and is granted access, but also who has
access to the bits and bytes in transit, or how to protect our data’s con-
fidentiality should a remote system be compromised. Is transport security
assumed, or do we rely on secure network connections (for example on the
network layer via a VPN or IPsec tunnels, or on the application layer via
TLS or SSH, to name but a few examples)? Is data stored “in the cloud”
accessible to anybody with access to the physical device (simple, but trust
is placed into the storage provider, often a third party), or do we encrypt it
(more secure, but also more complex)? Under what jurisdiction is the data
stored, and what data could be subpoenaed by which authorities? Would
you even be notified if access to your and your users’ data was granted to
law enforcement?
As we will see throughout this book, there never is a single solution to

all our problems. We will continually have to make trade-o↵s based on as-
sumptions and probabilities, usability considerations and practical concerns.
There is no silver bullet, no one solution, no unique, always correct approach.
This is why it is so important for a system administrator to be aware of all
the factors, and to ask the questions we’ve started to suggest here at the end
of this chapter. Not surprisingly, this line of questions quickly leads us be-
yond the basic concepts of file systems and storage models, but likewise often
brings us full circle when we have to consider higher level applications and
consider just how exactly they access data and what kinds of assumptions
and requirements they pose as to availability, security and performance. The
topics covered in this chapter are the foundation upon which our systems are
built.
Problems and Exercises
Problems
1. Identify the storage area model(s) predominantly used in your envi-
ronment(s). What kind of problems with each do you frequently en-
counter? Would changing the storage model be a feasible solution to
these problems? Why or why not?
2. Research the implementation details of a popular cloud storage provider.

What storage solutions did they choose? Can you identify all the dif-
ferent models we discussed?
3. Identify at least three distinct security concerns in each of the storage

models we discussed. Outline in detail how an attacker might exploit
such weaknesses and how a System Adminstrator might prevent or
counter each attack.
4. Ask your system administrators if they have any old or broken hard
drives, if possible from di↵erent manufacturers or with di↵erent capac-
ities. Open up the drives and identify the various components. How
many read-write heads are there? How many platters? How do the
di↵erent models di↵er?
5. Compare the output of the fdisk(8) command on di↵erent operat-

ing systems. Divide a disk into a number of di↵erent partitions, then
use disklabel(8) to review the resulting label. Identify all the di↵er-
ent partitions, verify their capacities and explain any discrepancies or
unexpected findings.
6. Compare the various composite RAID levels and analyze their respec-
tive fault tolerance and mean time to recovery in the case of one or
122
more failed disks.

7. Assuming 500 GB disk drives, what is the total capacity of a RAID
configuration using 7 disks under RAID 0, RAID 1, RAID 5, RAID 6,
RAID 0+1, RAID 1+0 and RAID 5+0?
8. You want to create a RAID 5 using several multi-terabyte drives. Re-
search and choose your preferred hardware, paying particular attention
to their Unrecoverable Read Error rates (i.e. the mean time to failure,
usually expressed as number of bits read). If the array is filled to ca-
pacity with data, and one of the drives failes and has to be replaced,
how much data do you end up reading when you rebuild the array?
What does this say about the probability of an additional disk failure
while you’re still in degraded mode?
9. Identify the di↵erent file systems in use on the Unix versions you have
access to. How many partitions are there, where are they mounted,
what mount options do they use? Were they chosen explicitly or do
they appear to be OS defaults?
10. Identify all resources made available by the procfs system, both about
the system and about individual processes. Suppose procfs was not
available, how would you go about finding out the same information?
11. Identify the di↵erent devices found under /dev. What is the di↵er-
ence between /dev/null and /dev/zero, between /dev/random and
/dev/urandom?
12. Identify the file system limitations on a system you have access to.
What is the maximum file system size? What is the maximum file
size? What characters may be used in a file name? How many files can
you create in a single directory?
13. Compare at least four di↵erent file systems with respect to their default
block size and inode density. Consider both small (less than 1 GB in
size), “normal”, and very large (more than 1 TB in size) partitions or
volumes. How efficiently does each system manage this space? Can
you specify di↵erent inode and block sizes to increase efficiency?
14. Review the concepts of inode mappings to file names, hard links, and
symbolic links.
(a) Create a new file, then create a second hard link for this file. Verify
that both files are completely identical by using ls(1), stat(1)
and by appending data to one of the files and reading from the
other.
(b) Rename the original file and repeat – what changed? Why?
(c) Create a new file, then create a symbolic link for this file. Verify
that both the original file and the symbolic link are unique by
inspecting their inode numbers. Then append data to the file using
the regular name and confirm that reading from the symbolic link
yields the same data.
(d) Rename the original file and repeat – what changed? Why?
15. Create a very large file. Measure how long it takes to rename the file
within one directory using the mv(1) command. Next, use mv(1) to
move the file into a directory on a di↵erent file system or partition.
What do you observe? Explain the di↵erence.
16. Experiment with the usability, resilience and limitations of a system’s

root file system.
(a) Identify how much free disk space you have, then fill it up. Can
you use up all available space using a single file? What are the
e↵ects of using up all disk space? Can you still create new, empty
files? Can you still log in? Why/why not?
(b) Identify how many free inodes are available on your root file sys-
tem, then use them up by, e.g., creating lots and lots of empty files.
What happens to the directory size in which you create these files?
What is the error message once you run out of inodes? Can you
still create new files? Can you still write to existing files? How
much disk space is available now? Can you still log in? Why/why
not?
BIBLIOGRAPHY 125
Bibliography
[1] EMC Education Service, Information Storage and Management: Stor-
ing, Managing, and Protecting Digital Information, John Wiley & Sons,
2009
[2] Æleen Frisch, Essential System Administration, O’Reilly Media, 2002
[3] Sanjay Ghemawat, Howard Gobio↵, and Shun-Tak Leung, The Google
File System, in “Proceedings of the nineteenth ACM symposium on Op-
erating systems principles”, 2003, ACM, New York, NY; also available
on the Internet at http://research.google.com/archive/gfs-sosp2003.pdf
(visited September 6th, 2012)
[4] The Hadoop Distributed File System: Ar-

chitecture and Design, on the Internet at
https://hadoop.apache.org/common/docs/r0.14.4/hdfs design.html
(visited September 6th, 2012)
[5] Zooko Wilcox-O’Hearn, Brian Warner, Tahoe – The Least-Authority

Filesystem, in “Proceedings of the 4th ACM international workshop on
Storage security and survivability”, 2008, ACM, New York, NY; also
available on the Internet at http://tahoe-lafs.org/˜zooko/lafs.pdf (vis-
ited September 6th, 2012)
[6] Eric Steven Raymond, The Art of Unix Programming, Addison-

Wesley Professional, September 2003; also available on the Internet at
http://catb.org/˜esr/writings/taoup/ (visited January 7, 2012)
[7] W. Richard Stevens, Stephen A. Rago, Advanced Programming in the

UNIX Environment, Addison-Wesley Professional; 2nd edition, 2005
[8] Abram Silberschatz, Greg Gagne, Peter B. Galvin Operating System

Concepts, John Wiley & Sons, 2011
[9] Marshall Kirk McKusick, William N. Joy, Samuel J. Le✏er and Robert
S. Fabry A Fast File System for UNIX, ACM Transactions on Computer
Systems 2 (3), 1984, ACM, New York, NY; also available on the In-
ternet at http://www.cs.berkeley.edu/˜brewer/cs262/FFS.pdf (visited
September 9, 2012)

Of File Systems and Storage Models

Uploaded by

Copyright:

Available Formats

Of File Systems and Storage Models

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Of File Systems and Storage Models

Uploaded by

Copyright:

Available Formats

Chapter 4

Of File Systems and Storage

As system administrators, we are responsible for all kinds of devices: we

4.2 Storage Models

4.2.1 Direct Attached Storage

(a) Diagram (b) A Maxtor IDE Drive

Figure 4.1: Direct Attached Storage

Alternatively, multiple direct attached disks can be combined to create a

4.2.2 Network Attached Storage

Figure 4.2: Three hosts using Network Attached Storage, or NAS

In contrast to DAS, a dedicated file server generally contains significantly

4.2.3 Storage Area Networks

(a) Huawei Tecal RH 2288H V2 (b) NetApp FAS 3050c

Figure 4.3: NAS and SAN Hardware

storage size, data availability, data redundancy, and performance, it becomes

di↵erent consumers in a SAN to be able to independently address the distinct

4.2.4 Cloud Storage

as Amazon’s Simple Storage Service (S3), or Windows Azure’s Blob Storage;

applications and entire OS images can easily be deployed, imported or ex-

4.2.5 Storage Model Considerations

(a) Open Hard Disc Drive (b) SSD Mini PCIe

4.3 Disk Devices and Interfaces

lower power consumption and generally higher performance, the dominant

Likewise, a large variety of cables, connectors and adapters help illustrate

4.3.1 Physical Disk Structure

(a) Disk Structure: Cylinders/- (b) Physical Disk Structure with

True Hard Drive Capacity

As a result, a drive labeled as having a capacity of 500 GB will in fact

4.4 Dividing and Combining Disks

Device Boot Start End Blocks Id System

Listing 4.1: fdisk(8) sample invocation and output on a Linux system

Disk and BIOS Limitations

1024(cylinders) ⇤ 16(heads) ⇤ 63(sectors) ⇤ 512(bytes/sector) = 504M B

Enhanced BIOSes removed some of these limitations, allowing for

partitioning scheme including a small /boot partition created at the be-

To overcome the limitations of CHS addressing, data blocks on the

In 2012, the current ATA/ATAPI specification uses 48 bits, limiting disk

Listing 4.2: disklabel(8) invocation and output on a NetBSD system

is actually a virtual disk, the information reported relating to the hardware,

4.4.2 Logical Volumes

LVMs manage and combine three distinct resources, as illustrated in Fig-

By creating logical extents as storage units on the logical volumes, the

ment, a RAID configuration cannot be expanded or shrunk without data

RAID 0 Striped array (block level), no parity or mirroring

Table 4.1: A brief summary of standard RAID levels.

(a) Block-level striping (b) Mirroring

(c) Block-level striping with distributed parity

Composite RAID and Failure Rates

RAID 0+1 Mirrored array of stripes

Table 4.2: Standard RAID levels can be combined.

Spectacular File System Confusion

I immediately stopped the data transfer and even physically dis-

4.5 File Systems

4.5.1 File System Types

Disk File Systems

Distributed File Systems

“Other” File Systems

4.6 File System Layout