doc/venti/venti.html

<!doctype html public "-//W3C//DTD HTML 4.0 Transitional//EN">
<html>

<head>

<meta http-equiv="Content-Language" content="en-us">
<meta http-equiv="Content-Type" content="text/html; charset=iso-8859-1">

<title>Venti: a new approach to archival storage</title>
</head>

<body bgcolor="white">

<h1>Venti: a new approach to archival storage</h1>

<p>

Sean Quinlan and Sean Dorward
<br>

Bell Labs, Lucent Technologies
<p>

<h1>Abstract</h1>
<p>

This paper describes a network storage system, called Venti, intended
for archival data.  In this system, a unique hash of a block's
contents acts as the block identifier for read and write operations.
This approach enforces a write-once policy, preventing accidental or
malicious destruction of data.  In addition, duplicate copies of a
block can be coalesced, reducing the consumption of storage and
simplifying the implementation of clients.  Venti is a building block
for constructing a variety of storage applications such as logical
backup, physical backup, and snapshot file systems.
<p>

We have built a prototype of the system and present some preliminary
performance results.  The system uses magnetic disks as the storage
technology, resulting in an access time for archival data that is
comparable to non-archival data.  The feasibility of the write-once
model for storage is demonstrated using data from over a decade's use
of two Plan 9 file systems.
<p>

<h1>1.  Introduction</h1>
<p>

Archival storage is a second class citizen.  Many computer
environments provide access to a few recent versions of the
information stored in file systems and databases, though this access
can be tedious and may require the assistance of a system
administrator.  Less common is the ability for a user to examine data
from last month or last year or last decade.  Such a feature may not
be needed frequently, but when it is needed it is often crucial.
<p>

The growth in capacity of storage technologies exceeds the ability of
many users to generate data, making it practical to archive data in
perpetuity.  Plan 9, the computing environment that the authors use,
includes a file system that stores archival data to an optical jukebox
[16, 17].  Ken Thompson observed that, for our usage patterns, the
capacity of the jukebox could be considered infinite.  In the time it
took for us to fill the jukebox, the improvement in technology would
allow us to upgrade to a new jukebox with twice the capacity.
<p>

Abundant storage suggests that an archival system impose a write-once
policy.  Such a policy prohibits either a user or administrator from
deleting or modifying data once it is stored.  This approach greatly
reduces the opportunities for accidental or malicious data loss and
simplifies the system's implementation.
<p>

Moreover, our experience with Plan 9 is that a write-once policy
changes the way one views storage.  Obviously, some data is temporary,
derivative, or so large that it is either undesirable or impractical
to retain forever and should not be archived.  However, once it is
decided that the data is worth keeping, the resources needed to store
the data have been consumed and cannot be reclaimed.  This eliminates
the task of periodically "cleaning up" and deciding whether the data
is still worth keeping.  More thought is required before storing the
data to a write-once archive, but as the cost of storage continues to
fall, this becomes an easy decision.
<p>

This paper describes the design and implementation of an archival
server, called Venti.  The goal of Venti is to provide a write-once
archival repository that can be shared by multiple client machines and
applications.  In addition, by using magnetic disks as the primary
storage technology, the performance of the system approaches that of
non-archival storage.
<p>

<h1>2.  Background</h1>
<p>

A prevalent form of archival storage is the regular backup of data to
magnetic tape [15].  A typical scenario is to provide backup as a
central service for a number of client machines.  Client software
interfaces with a database or file system and determines what data to
back up.  The data is copied from the client to the tape device, often
over a network, and a record of what was copied is stored in a catalog
database.
<p>

Restoring data from a tape backup system can be tedious and error
prone.  The backup system violates the access permission of the file
system, requiring a system administrator or privileged software to
perform the task.  Since they are tedious, restore operations are
infrequent and problems with the process may go undetected.  Potential
sources of error abound: tapes are mislabeled or reused or lost,
drives wander out of alignment and cannot read their old tapes,
technology becomes obsolete.
<p>

For tape backup systems, a tradeoff exists between the performance of
backup and restore operations [1].  A full backup simplifies the
process of restoring data since all the data is copied to a continuous
region on the tape media.  For large file systems and databases,
incremental backups are more efficient to generate, but such backups
are not self-contained; the data for a restore operation is scattered
across multiple incremental backups and perhaps multiple tapes.  The
conventional solution is to limit the extent of this scattering by
performing a full backup followed by a small number of incremental
backups.
<p>

File systems such as Plan 9 [16, 17], WAFL [5], and AFS [7] provide a
more unified approach to the backup problem by implementing a snapshot
feature.  A snapshot is a consistent read-only view of the file system
at some point in the past.  The snapshot retains the file system
permissions and can be accessed with standard tools (ls, cat, cp,
grep, diff) without special privileges or assistance from an
administrator.  In our experience, snapshots are a relied-upon and
frequently-used resource because they are always available and easy to
access.
<p>

Snapshots avoid the tradeoff between full and incremental backups.
Each snapshot is a complete file system tree, much like a full backup.
The implementation, however, resembles an incremental backup because
the snapshots and the active file system share any blocks that remain
unmodified; a snapshot only requires additional storage for the blocks
that have changed.  To achieve reasonable performance, the device that
stores the snapshots must efficiently support random access, limiting
the suitability of tape storage for this approach.
<p>

In the WAFL and AFS systems, snapshots are ephemeral; only a small
number of recent versions of the file system are retained.  This
policy is reasonable since the most recent versions of files are the
most useful.  For these systems, archival storage requires an
additional mechanism such as tape backup.
<p>

The philosophy of the Plan 9 file system is that random access storage
is sufficiently cheap that it is feasible to retain snapshots
permanently.  The storage required to retain all daily snapshots of a
file system is surprisingly modest; later in the paper we present
statistics for two file servers that have been in use over the last 10
years.
<p>

Like Plan 9, the Elephant file system [18] retains many versions of
data.  This system allows a variety of storage reclamation policies
that determine when a version of a file should be deleted.  In
particular, "landmark" versions of files are retained permanently and
provide an archival record.
<p>

<h1>3.  The Venti Archival Server</h1>
<p>

Venti is a block-level network storage system intended for archival
data.  The interface to the system is a simple protocol that enables
client applications to read and write variable sized blocks of data.
Venti itself does not provide the services of a file or backup system,
but rather the backend archival storage for these types of
applications.
<p>

Venti identifies data blocks by a hash of their contents.  By using a
collision-resistant hash function with a sufficiently large output, it
is possible to consider the hash of a data block as unique.  Such a
unique hash is called the fingerprint of a block and can be used as
the address for read and write operations.  This approach results in a
storage system with a number of interesting properties.
<p>

As blocks are addressed by the fingerprint of their contents, a block
cannot be modified without changing its address; the behavior is
intrinsically write-once.  This property distinguishes Venti from most
other storage systems, in which the address of a block and its
contents are independent.
<p>

Moreover, writes are idempotent.  Multiple writes of the same data can
be coalesced and do not require additional storage space.  This
property can greatly increase the effective storage capacity of the
server since it does not rely on the behavior of client applications.
For example, an incremental backup application may not be able to
determine exactly which blocks have changed, resulting in unnecessary
duplication of data.  On Venti, such duplicate blocks will be
discarded and only one copy of the data will be retained.  In fact,
replacing the incremental backup with a full backup will consume the
same amount of storage.  Even duplicate data from different
applications and machines can be eliminated if the clients write the
data using the same block size and alignment.
<p>

The hash function can be viewed as generating a universal name space
for data blocks.  Without cooperating or coordinating, multiple
clients can share this name space and share a Venti server.  Moreover,
the block level interface places few restrictions on the structures
and format that clients use to store their data.  In contrast,
traditional backup and archival systems require more centralized
control.  For example, backup systems include some form of job
scheduler to serialize access to tape devices and may only support a
small number of predetermined data formats so that the catalog system
can extract pertinent meta-data.
<p>

Venti provides inherent integrity checking of data.  When a block is
retrieved, both the client and the server can compute the fingerprint
of the data and compare it to the requested fingerprint.  This
operation allows the client to avoid errors from undetected data
corruption and enables the server to identify when error recovery is
necessary.
<p>

Using the fingerprint of a block as its identity facilitates features
such as replication, caching, and load balancing.  Since the contents
of a particular block are immutable, the problem of data coherency is
greatly reduced; a cache or a mirror cannot contain a stale or out of
date version of a block.
<p>

<h2>3.1.  Choice of Hash Function</h2>
<p>

The design of Venti requires a hash function that generates a unique
fingerprint for every data block that a client may want to store.
Obviously, if the size of the fingerprint is smaller than the size of
the data blocks, such a hash function cannot exist since there are
fewer possible fingerprints than blocks.  If the fingerprint is large
enough and randomly distributed, this problem does not arise in
practice.  For a server of a given capacity, the likelihood that two
different blocks will have the same hash value, also known as a
collision, can be determined.  If the probability of a collision is
vanishingly small, we can be confident that each fingerprint is
unique.
<p>

It is desirable that Venti employ a cryptographic hash function.  For
such a function, it is computationally infeasible to find two distinct
inputs that hash to the same value [10].  This property is important
because it prevents a malicious client from intentionally creating
blocks that violate the assumption that each block has a unique
fingerprint.  As an additional benefit, using a cryptographic hash
function strengthens a client's integrity check, preventing a
malicious server from fulfilling a read request with fraudulent data.
If the fingerprint of the returned block matches the requested
fingerprint, the client can be confident the server returned the
original data.
<p>

Venti uses the Sha1 hash function [13] developed by the US National
Institute for Standards and Technology (NIST).  Sha1 is a popular hash
algorithm for many security systems and, to date, there are no known
collisions.  The output of Sha1 is a 160 bit (20 byte) hash value.
Software implementations of Sha1 are relatively efficient; for
example, a 700Mhz Pentium 3 can compute the Sha1 hash of 8 Kbyte data
blocks in about 130 microseconds, a rate of 60 Mbytes per second.
<p>

Are the 160 bit hash values generated by Sha1 large enough to ensure
the fingerprint of every block is unique?  Assuming random hash values
with a uniform distribution, a collection of n different data blocks
and a hash function that generates b bits, the probability p that
there will be one or more collisions is bounded by the number of pairs
of blocks multiplied by the probability that a given pair will
collide, i.e.
<p>
&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
<img src="probablity.gif" ALT="probablity">
<p>

Today, a large storage system may contain a petabyte (10^15 bytes) of data.
Consider an even larger system that contains an exabyte (10^18 bytes)
stored as 8 Kbyte blocks (~10^14 blocks).  Using the Sha1 hash function, the
probability of a collision is less than 10^-20.  Such a scenario seems
sufficiently unlikely that we ignore it and use the Sha1 hash as a
unique identifier for a block.  Obviously, as storage technology
advances, it may become feasible to store much more than an exabyte,
at which point it maybe necessary to move to a larger hash function.
NIST has already proposed variants of Sha1 that produce 256, 384, and
512 bit results [14].  For the immediate future, however, Sha1 is a
suitable choice for generating the fingerprint of a block.
<p>

<h2>3.2.  Choice of Storage Technology</h2>
<p>

When the Plan 9 file system was designed in 1989, optical jukeboxes
offered high capacity with respectable random access performance and
thus were an obvious candidate for archival storage.  The last decade,
however, has seen the capacity of magnetic disks increase at a far
faster rate than optical technologies [20].  Today, a disk array costs
less than the equivalent capacity optical jukebox and occupies less
physical space.  Disk technology is even approaching tape in cost per
bit.
<p>

Magnetic disk storage is not as stable or permanent as optical media.
Reliability can be improved with technology such as RAID, but unlike
write-once optical disks, there is little protection from erasure due
to failures of the storage server or RAID array firmware.  This issue
is discussed in Section 7.
<p>

Using magnetic disks for Venti has the benefit of reducing the
disparity in performance between conventional and archival storage.
Operations that previously required data to be restored to magnetic
disk can be accomplished directly from the archive.  Similarly, the
archive can contain the primary copy of often-accessed read-only data.
In effect, archival data need not be further down the storage
hierarchy; it is differentiated by the write-once policy of the
server.
<p>

<h1>4.  Applications</h1>
<p>

Venti is a building block on which to construct a variety of storage
applications.  Venti provides a large repository for data that can be
shared by many clients, much as tape libraries are currently the
foundation of many centralized backup systems.  Applications need to
accommodate the unique properties of Venti, which are different from
traditional block level storage devices, but these properties enable a
number of interesting features.
<p>

Applications use the block level service provided by Venti to store
more complex data structures.  Data is divided into blocks and written
to the server.  To enable this data to be retrieved, the application
must record the fingerprints of these blocks.  One approach is to pack
the fingerprints into additional blocks, called pointer blocks, that
are also written to the server, a process that can be repeated
recursively until a single fingerprint is obtained.  This fingerprint
represents the root of a tree of blocks and corresponds to a
hierarchical hash of the original data.
<p>

A simple data structure for storing a linear sequence of data blocks
is shown in Figure 1.  The data blocks are located via a fixed depth
tree of pointer blocks which itself is addressed by a root
fingerprint.  Applications can use such a structure to store a single
file or to mimic the behavior of a physical device such as a tape or a
disk drive.  The write-once nature of Venti does not allow such a tree
to be modified, but new versions of the tree can be generated
efficiently by storing the new or modified data blocks and reusing the
unchanged sections of the tree as depicted in Figure 2.
<p>


&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
<img src="SimpleTree.gif" ALT="simple tree">
<p>
Figure 1.  A tree structure for storing a linear sequence of blocks
<p>


&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
<img src="ModifiedTree.gif" ALT="modified tree">
<p>
Figure 2.  Build a new version of the tree.
<p>

By mixing data and fingerprints in a block, more complex data
structures can be constructed.  For example, a structure for storing a
file system may include three types of blocks: directory, pointer, and
data.  A directory block combines the meta information for a file and
the fingerprint to a tree of data blocks containing the file's
contents.  The depth of the tree can be determined from the size of
the file, assuming the pointer and data blocks have a fixed size.
Other structures are obviously possible.  Venti's block-level
interface leaves the choice of format to client applications and
different data structures can coexist on a single server.
<p>

The following sections describes three applications that use Venti as
an archival data repository: a user level archive utility called vac,
a proposal for a physical level backup utility, and our preliminary
work on a new version of the Plan 9 file system.
<p>

<h2>4.1.  Vac</h2>
<p>

Vac is an application for storing a collection of files and
directories as a single object, similar in functionality to the
utilities tar and zip.  With vac, the contents of the selected files
are stored as a tree of blocks on a Venti server.  The root
fingerprint for this tree is written to a vac archive file specified
by the user, which consists of an ASCII representation of the 20 byte
root fingerprint plus a fixed header string, and is always 45 bytes
long.  A corresponding program, called unvac, enables the user to
restore files from a vac archive.  Naturally, unvac requires access to
the Venti server that contains the actual data, but in most situations
this is transparent.  For a user, it appears that vac compresses any
amount of data down to 45 bytes.
<p>

An important attribute of vac is that it writes each file as a
separate collection of Venti blocks, thus ensuring that duplicate
copies of a file will be coalesced on the server.  If multiple users
vac the same data, only one copy will be stored on the server.
Similarly, a user may repeatedly vac a directory over time and even if
the contents of the directory change, the additional storage consumed
on the server will be related to the extent of the changes rather than
the total size of the contents.  Since Venti coalesces data at the
block level, even files that change may share many blocks with
previous versions and thus require little space on the server; log and
database files are good examples of this scenario.
<p>

On many Unix systems, the dump utility is used to back up file
systems.  Dump has the ability to perform incremental backups of data;
a user specifies a dump level, and only files that are new or have
changed since the last dump at this level are written to the archive.
To implement incremental backups, dump examines the modified time
associated with each file, which is an efficient method of filtering
out the unchanged files.
<p>

Vac also implements an incremental option based on the file
modification times.  The user specifies an existing vac file and this
archive is used to reduce the number of blocks written to the Venti
server.  For each file, vac examines the modified time in both the
file system and the vac archive.  If they are the same, vac copies the
fingerprint for the file from the old archive into the new archive.
Copying just the 20-byte fingerprint enables the new archive to
include the entire file without reading the data from the file system
nor writing the data across the network to the Venti server.  In
addition, unlike an incremental dump, the resulting archive will be
identical to an archive generated without the incremental option; it
is only a performance improvement.  This means there is no need to
have multiple levels of backups, some incremental, some full, and so
restore operations are greatly simplified.
<p>

A variant of the incremental option improves the backup of files
without reference to modification times.  As vac reads a file, it
computes the fingerprint for each block.  Concurrently, the pointer
blocks of the old archive are examined to determine the fingerprint
for the block at the same offset in the old version of the file.  If
the fingerprints are the same, the block does not need to be written
to Venti.  Instead, the fingerprint can simply be copied into the
appropriate pointer block.  This optimization reduces the number of
writes to the Venti server, saving both network and disk bandwidth.
Like the file level optimization above, the resulting vac file is no
different from the one produced without this optimization.  It does,
however, require the data for the file to be read and is only
effective if there are a significant number of unchanged blocks.
<p>

<h2>4.2.  Physical backup</h2>
<p>

Utilities such as vac, tar, and dump archive data at the file or
logical level: they walk the file hierarchy converting both data and
meta-data into their own internal format.  An alternative approach is
block-level or physical backup, in which the disk blocks that make up
the file system are directly copied without interpretation.  Physical
backup has a number of benefits including simplicity and potentially
much higher throughput [8].  A physical backup utility for file
systems that stores the resulting data on Venti appears attractive,
though we have not yet implemented such an application.
<p>

The simplest form of physical backup is to copy the raw contents of
one or mores disk drives to Venti.  The backup also includes a tree of
pointer blocks, which enables access to the data blocks.  Like vac,
the end result is a single fingerprint representing the root of the
tree; that fingerprint needs to be recorded outside of Venti.
<p>

Coalescing duplicate blocks is the main advantage of making a physical
backup to Venti rather than copying the data to another storage medium
such as tape.  Since file systems are inherently block based, we
expect coalescing to be effective.  Not only will backups of a file
system over time share many unchanged blocks, but even file systems
for different machines that are running the same operating system may
have many blocks in common.  As with vac, the user sees a full backup
of the device, while retaining the storage space advantages of an
incremental backup.
<p>

One enhancement to physical backup is to copy only blocks that are
actively in use in the file system.  For most file system formats it
is relatively easy to determine if a block is in use or free without
walking the file system hierarchy.  Free blocks generally contain the
remnants of temporary files that were created and removed in the time
between backups and it is advantageous not to store such blocks.  This
optimization requires that the backup format be able to represent
missing blocks, which can easily be achieved on Venti by storing a
null value for the appropriate entry in the pointer tree.
<p>

The random access performance of Venti is sufficiently good that it is
possible to use a physical backup without first restoring it to disk.
With operating system support, it is feasible to directly mount a
backup file system image from Venti.  Access to this file system is
read only, but it provides a natural method of restoring a subset of
files.  For situations where a full restore is required, it might be
possible to do this restore in a lazy fashion, copying blocks from
Venti to the file system as needed, instead of copying the entire
contents of the file system before resuming normal operation.
<p>

The time to perform a physical backup can be reduced using a variety
of incremental techniques.  Like vac, the backup utility can compute
the fingerprint of each block and compare this fingerprint with the
appropriate entry in the pointer tree of a previous backup.  This
optimization reduces the number of writes to the Venti server.  If the
file system provides information about which blocks have changed, as
is the case with WAFL, the backup utility can avoid even reading the
unchanged blocks.  Again, a major advantage of using Venti is that the
backup utility can implement these incremental techniques while still
providing the user with a full backup.  The backup utility writes the
new blocks to the Venti server and constructs a pointer tree with the
appropriate fingerprint for the unchanged blocks.
<p>

<h2>4.3.  Plan 9 File system</h2>
<p>

When combined with a small amount of read/write storage, Venti can be
used as the primary location for data rather than a place to store
backups.  A new version of the Plan 9 file system, which we are
developing, exemplifies this approach.
<p>

Previously, the Plan 9 file system was stored on a combination of
magnetic disks and a write-once optical jukebox.  The jukebox
furnishes the permanent storage for the system, while the magnetic
disks act as a cache for the jukebox.  The cache provides faster file
access and, more importantly, accumulates the changes to the file
system during the period between snapshots.  When a snapshot is taken,
new or modified blocks are written from the disk cache to the jukebox.
<p>

The disk cache can be smaller than the active file system, needing
only to be big enough to contain the daily changes to the file system.
However, accesses that miss the cache are significantly slower since
changing platters in the jukebox takes several seconds.  This
performance penalty makes certain operations on old snapshots
prohibitively expensive.  Also, on the rare occasions when the disk
cache has been reinitialized due to corruption, the file server spends
several days filling the cache before performance returns to normal.
<p>

The new version of the Plan 9 file system uses Venti instead of an
optical jukebox as its storage device.  Since the performance of Venti
is comparable to disk, this substitution equalizes access both to the
active and to the archival view of the file system.  It also allows
the disk cache to be quite small; the cache accumulates changes to the
file system between snapshots, but does not speed file access.
<p>

<h1>5.  Implementation</h1>
<p>

We have implemented a prototype of Venti.  The implementation uses an
append-only log of data blocks and an index that maps fingerprints to
locations in this log.  It also includes a number of features that
improve robustness and performance.  This section gives a brief
overview of the implementation.  Figure 3 shows a block diagram of the
server.
<p>


&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
<img src="Block.gif" ALT="block diagram">
<p>
Figure 3.  A block diagram of the Venti prototype.
<p>

Since Venti is intended for archival storage, one goal of our
prototype is robustness.  The approach we have taken is to separate
the storage of data blocks from the index used to locate a block.  In
particular, blocks are stored in an append-only log on a RAID array of
disk drives.  The simplicity of the append-only log structure
eliminates many possible software errors that might cause data
corruption and facilitates a variety of additional integrity
strategies.  A separate index structure allows a block to be
efficiently located in the log; however, the index can be regenerated
from the data log if required and thus does not have the same
reliability constraints as the log itself.
<p>

The structure of the data log is illustrated in Figure 4.  To ease
maintenance, the log is divided into self-contained sections called
arenas.  Each arena contains a large number of data blocks and is
sized to facilitate operations such as copying to removable media.
Within an arena is a section for data bocks that is filled in an
append-only manner.  In Venti, data blocks are variable sized, up to a
current limit of 52 Kbytes, but since blocks are immutable they can be
densely packed into an arena without fragmentation.
<p>

&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
<img src="LogFormat.gif" ALT="log format">
<p>
Figure 4.  The format of the data log.
<p>

Each block is prefixed by a header that describes the contents of the
block.  The primary purpose of the header is to provide integrity
checking during normal operation and to assist in data recovery.  The
header includes a magic number, the fingerprint and size of the block,
the time when the block was first written, and identity of the user
that wrote it.  The header also includes a user-supplied type
identifier, which is explained in Section 7.  Note, only one copy of a
given block is stored in the log, thus the user and wtime fields
correspond to the first time the block was stored to the server.
<p>

Before storing a block in the log, an attempt is made to compress its
contents.  The inclusion of data compression increases the effective
capacity of the archive and is simple to add given the log structure.
Obviously, some blocks are incompressible.  The encoding field in the
block header indicates whether the data was compressed and, if so, the
algorithm used.  The esize field indicates the size of the data after
compression, enabling the location of the next block in the arena to
be determined.  The downside of using compression is the computational
cost, typically resulting in a decrease in the rate that blocks can be
stored and retrieved.  Our prototype uses a custom Lempel-Ziv '77 [21]
algorithm that is optimized for speed.  Compression is not a
performance bottleneck for our existing server.  Future
implementations may benefit from hardware solutions.
<p>

In addition to a log of data blocks, an arena includes a header, a
directory, and a trailer.  The header identifies the arena.  The
directory contains a copy of the block header and offset for every
block in the arena.  By replicating the headers of all the blocks in
one relatively small part of the arena, the server can rapidly check
or rebuild the system's global block index.  The directory also
facilitates error recovery if part of the arena is destroyed or
corrupted.  The trailer summarizes the current state of the arena
itself, including the number of blocks and the size of the log.
Within the arena, the data log and the directory start at opposite
ends and grow towards each other.  When the arena is filled, it is
marked as sealed, and a fingerprint is computed for the contents of
the entire arena.  Sealed arenas are never modified.
<p>

The basic operation of Venti is to store and retrieve blocks based on
their fingerprints.  A fingerprint is 160 bits long, and the number of
possible fingerprints far exceeds the number of blocks stored on a
server.  The disparity between the number of fingerprints and blocks
means it is impractical to map the fingerprint directly to a location
on a storage device.  Instead, we use an index to locate a block
within the log.
<p>

We implement the index using a disk-resident hash table as illustrated
in Figure 5.  The index is divided into fixed-sized buckets, each of
which is stored as a single disk block.  Each bucket contains the
index map for a small section of the fingerprint space.  A hash
function is used to map fingerprints to index buckets in a roughly
uniform manner, and then the bucket is examined using binary search.
If provisioned with sufficient buckets, the index hash table will be
relatively empty and bucket overflows will be extremely rare.  If a
bucket does overflow, the extra entries are placed in an adjacent
bucket.  This structure is simple and efficient, requiring one disk
access to locate a block in almost all cases.
<p>


<p>
&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
<img src="Index.gif" ALT="index format">
<p>

Figure 5.  Format of the index.
<p>

The need to go through an index is the main performance penalty for
Venti compared to a conventional block storage device.  Our prototype
uses three techniques to increase the performance: caching, striping,
and write buffering.
<p>

The current implementation has two important caches of approximately
equal size: a block cache and an index cache.  A hit in the block
cache returns the data for that fingerprint, bypassing the both the
index lookup and access to the data log.  Hits in the index cache
eliminate only the index lookup, but the entries are much smaller and
the hit rate correspondingly higher.
<p>

Unfortunately, these caches do not speed the process of storing a new
block to Venti.  The server must check that the block is not a
duplicate by examining the index.  If the block is not contained on
the server, it will obviously not be in any cache.  Since the
fingerprint of the block contains no internal structure, the location
of a fingerprint in the index is essentially random.  Furthermore, the
archival nature of Venti means the entire index will not fit in memory
because of the large number of blocks.  Combining these factors means
that the write performance of Venti will be limited to the random IO
performance of the index disk, which for current technology is a few
hundred accesses per second.  By striping the index across multiple
disks, however, we get a linear speedup.  This requires a sufficient
number of concurrent accesses, which we assure by buffering the writes
before accessing the index.
<p>

The prototype Venti server is implemented for the Plan 9 operating
system in about 10,000 lines of C. The server runs on a dedicated dual
550Mhz Pentium III processor system with 2 Gbyte of memory and is
accessed over a 100Mbs Ethernet network.  The data log is stored on a
500 Gbyte MaxTronic IDE Raid 5 Array and the index resides on a string
of 8 Seagate Cheetah 18XL 9 Gbyte SCSI drives.
<p>

<h1>6.  Performance</h1>
<p>

Table 1 gives the preliminary performance results for read and write
operations in a variety of situations.  For comparison, we include the
SCSI performance of the RAID array.  Although the performance is still
several times slower than directly accessing the disk, we believe the
results are promising and will improve as the system matures.
<p>
Table 1.  The performance of read and write operations in Mbytes/s for 8 Kbyte blocks.
<p>
<p>
<table align=center>
<tr>
<th></th>
<th width=150>sequential reads</th>
<th width=150>random reads</th>
<th width=150>virgin writes</th>
<th width=150>duplicate writes</th>
</tr>
<tr>
<td>uncached</td>
<td align=center>0.9</td>
<td align=center>0.4</td>
<td align=center>3.7</td>
<td align=center>5.6</td>
</tr>
<tr>
<td>index cache</td>
<td align=center>4.2</td>
<td align=center>0.7</td>
<td align=center>-</td>
<td align=center>6.2</td>
</tr>
<tr>
<td>block cache</td>
<td align=center>6.8</td>
<td align=center>-</td>
<td align=center>-</td>
<td align=center>6.5</td>
</tr>
<tr>
<td>raw raid</td>
<td align=center>14.8</td>
<td align=center>1.0</td>
<td align=center>12.4</td>
<td align=center>12.4</td>
</tr>
</table>
<p>


The uncached sequential read performance is particularly bad.  The
problem is that these sequential reads require a random read of the
index.  Without assistance from the client, the read operations are
not overlapped and do not benefit from the striping of the index.  One
possible solution is a form of read-ahead.  When reading a block from
the data log, it is feasible to also read several following blocks.
These extra blocks can be added to the caches without referencing the
index.  If blocks are read in the same order they were written to the
log, the latency of uncached index lookups will be avoided.  This
strategy should work well for streaming data such as multimedia files.
<p>

The basic assumption in Venti is that the growth in capacity of disks
combined with the removal of duplicate blocks and compression of their
contents enables a model in which it is not necessary to reclaim space
by deleting archival data.  To demonstrate why we believe this model
is practical, we present some statistics derived from a decade's use
of the Plan 9 file system.
<p>

The computing environment in which we work includes two Plan 9 file
servers named bootes and emelie.  Bootes was our primary file
repository from 1990 until 1997 at which point it was superseded by
emelie.  Over the life of these two file servers there have been 522
user accounts of which between 50 and 100 were active at any given
time.  The file servers have hosted numerous development projects and
also contain several large data sets including chess end games,
astronomical data, satellite imagery, and multimedia files.
<p>

Figure 6 depicts the size of the active file system as measured over
time by du, the space consumed on the jukebox, and the size of the
jukebox's data if it were to be stored on Venti.  The ratio of the
size of the archival data and the active file system is also given.
As can be seen, even without using Venti, the storage required to
implement the daily snapshots in Plan 9 is relatively modest, a result
of the block level incremental approach to generating a snapshot.
When the archival data is stored to Venti the cost of retaining the
snapshots is reduced significantly.  In the case of the emelie file
system, the size on Venti is only slightly larger than the active file
system; the cost of retaining the daily snapshots is almost zero.
Note that the amount of storage that Venti uses for the snapshots
would be the same even if more conventional methods were used to back
up the file system.  The Plan 9 approach to snapshots is not a
necessity, since Venti will remove duplicate blocks.
<p>
<img src="bootes.gif" ALT="storage sizes for bootes">
<img src="emelie.gif" ALT="storage sizes for emelie">
<img src="bootes2.gif" ALT="ratio of sizes for bootes">
<img src="emelie2.gif" ALT="ratio of sizes for emelie">
<p>
Figure 6. Graphs of the various sizes of two Plan 9 file servers.
<p>

When stored on Venti, the size of the jukebox data is reduced by three
factors: elimination of duplicate blocks, elimination of block
fragmentation, and compression of the block contents.  Table 2
presents the percent reduction for each of these factors.  Note,
bootes uses a 6 Kbyte block size while emelie uses 16 Kbyte, so the
effect of removing fragmentation is more significant on emelie.
<p>

The 10 year history of the two Plan 9 file servers may be of interest
to other researchers.  We have made available per-block information
including a hash of each block's contents, all the block pointers, and
most of the directory information.  The traces do not include the
actual contents of files nor the file names.  There is sufficient
information to reconstruct the structure of the file system and to
track the daily changes to this structure over time.  The traces are
available at http://www.cs.bell-labs.com/~seanq/p9trace.html.
<p>

Table 2.  The percentage reduction in the size of data stored on
Venti.
<p>
<table align=center>
<tr>
<th></th>
<th width=150>bootes</th>
<th width=150>emelie</th>
</tr>
<tr>
<td>Elimination of duplicates</td>
<td align=center>27.8%</td>
<td align=center>31.3%</td>
</tr>
<tr>
<td>Elimination of fragments</td>
<td align=center>10.2%</td>
<td align=center>25.4%</td>
</tr>
<tr>
<td>Data Compression</td>
<td align=center>33.8%</td>
<td align=center>54.1%</td>
</tr>
<tr>
<td>Total Reduction</td>
<td align=center>59.7%</td>
<td align=center>76.5%</td>
</tr>
</table>
<p>


<p>

<h1>7.  Reliability and Recovery</h1>
<p>

In concert with the development of the Venti prototype, we have built
a collection of tools for integrity checking and error recovery.
Example uses of these tools include: verifying the structure of an
arena, checking there is an index entry for every block in the data
log and vice versa, rebuilding the index from the data log, and
copying an arena to removable media.  These tools directly access the
storage devices containing the data log and index and are executed on
the server.
<p>

The directory structure at the end of each area enhances the
efficiency of many integrity and recovery operations, since it is
typically two orders of magnitude smaller than the arena, yet contains
most of the needed information.  The index checking utility, for
example, is implemented as a disk based sort of all the arena
directories, followed by a comparison between this sorted list and the
index.  Our prototype currently contains approximately 150 million
blocks using 250 Gbytes of storage.  An index check takes 2.2 hours,
which is significantly less than the 6 hours it takes to read all the
log data.
<p>

An additional integrity and recovery feature is the association of a
type identifier with every block.  This 8 bit identifier is included
with all client read and write operations and has the effect of
partitioning the server into multiple independent domains.  The idea
is that type indicates the interpretation of the data contained in the
block.  A client can use this feature, for example, to indicate that a
block is the root node for a tree of blocks.  Currently, the data
format associated with a type is left entirely to the client; the
server does not interpret the type other that to use it in conjunction
with a fingerprint as the key with which to index a block.
<p>

One use of the type identifier is to assist the administrator in
locating blocks for which a user has accidentally lost the
fingerprint.  Using a tool on the server, the data log can be scanned
for blocks that match specified criteria, including the block type,
the write time, and user identifier.  The type makes it relatively
simple to locate forgotten root blocks.  Future uses for the type
might include the ability for the server to determine the location of
fingerprints within a block, enabling the server to traverse the data
structures that have been stored.
<p>

By storing the data log on a RAID 5 disk array, our server is
protected against single drive failures.  Obviously, there are many
scenarios where this is not sufficient: multiple drives may fail,
there may be a fire in the machine room, the RAID firmware may contain
bugs, or the device may be stolen.
<p>

Additional protection could be obtained by using one or more off-site
mirrors for the server.  We have not yet implemented this strategy,
but the architecture of Venti makes this relatively simple.  A
background process on the server copies new blocks from the data log
to the mirrors.  This copying can be achieved using the Venti
protocol; the server is simply another client to the mirror.
<p>

Even mirroring may not be sufficient.  The implementation of Venti may
contain bugs that can be exploited to compromise the server.  An
automated attack may delete data on many servers simultaneously.
Storage devices that provide low level enforcement of a write-once
policy would provide protection for such an attack.  Write-once
read-many optical jukeboxes often provide such protection, but this is
not yet common for magnetic disk based storage systems.  We have thus
resorted to copying the sealed arenas onto removable media.
<p>

<h1>8.  Related Work</h1>
<p>

The Stanford Archival Vault [2] is a prototype archival repository
intended for digital libraries.  The archive consists of a write-once
log of digital objects (files) and several auxiliary indexes for
locating objects within the log.  Objects are identified by the hash
of their contents using a cyclic redundancy check (CRC).  Unlike
Venti, this system has no way to share data between objects that are
partially the same, or to build up complex data structures such as a
file system hierarchy.  Rather, the archive consists of a collection
of separate objects with a limited ability to group objects into sets.
<p>

On Venti, blocks are organized into more complex data structures by
creating hash-trees, an idea originally proposed by Merkle [11] for an
efficient digital signature scheme.
<p>

The approach to block retrieval in the Read-Only Secure File System
(SFSRO) [3] is comparable to Venti.  Blocks are identified by the Sha1
hash of their contents and this idea is applied recursively to build
up more complex structures.  The focus of this system is security, not
archival storage.  An administrator creates a digitally signed
database offline.  The database contains a public read-only file
system that can be published on multiple servers and efficiently and
securely accessed by clients.  SFSRO outperforms traditional methods
for providing data integrity between a client and a file server,
demonstrating an attractive property of hash-based addressing.
<p>

Given their similarities, it would be simple to implement SFSRO on top
of Venti.  The goal of Venti is to provide a flexible location for
archival storage and SFSRO is a good example of an application that
could use this capability.  In fact, using Venti would provide a
trivial solution to SFSRO's problem with stale NFS handles since data
is never deleted from Venti and thus a stale handle will never be
encountered.
<p>

Content-Derived Names [6] are another example of naming objects based
on a secure hash of its contents.  This work addresses the issue of
naming and managing the various binary software components, in
particular shared libraries, that make up an application.
<p>

The philosophy of the Elephant file system [18] is similar to Venti;
large, cheap disks make it feasible to retain many versions of data.
A feature of the Elephant system is the ability to specify a variety
of data retention policies, which can be applied to individual files
or directories.  These policies attempt to strike a balance between
the costs and benefits of storing every version of a file.  In
contrast, Venti focuses on the problem of how to store information
after deciding that it should be retained in perpetuity.  A system
such as the Elephant file system could incorporate Venti as the
storage device for the permanent "landmark" versions of files, much as
the Plan 9 file system will use Venti to archive snapshots.
<p>

Self-Securing Storage [19] retains all versions of file system data in
order to provide diagnosis and recovery from security breaches.  The
system is implemented as a self-contained network service that exports
an object-based disk interface, providing protection from compromise
of the client operating system.  Old data is retained for a window of
time and then deleted to reclaim storage.
<p>

Venti provides many of the features of self-securing storage: the
server is self-contained and accessed through a simple low-level
protocol, malicious users cannot corrupt or delete existing data on
the server, and old versions of data are available for inspection.  It
is unlikely that a system would write every file system operation to
Venti since storage is never reclaimed, but not deleting data removes
the constraint that an intrusion must be detected within a limited
window of time.  A hybrid approach might retain every version for some
time and some versions for all time.  Venti could provide the
long-term storage for such a hybrid.
<p>

<h1>9.  Future Work</h1>
<p>

Venti could be distributed across multiple machines; the approach of
identifying data by a hash of its contents simplifies such an
extension.  For example, the IO performance could be improved by
replicating the server and using a simple load balancing algorithm.
When storing or retrieving a block, clients direct the operation to a
server based on a few bits of the fingerprint.  Such load balancing
could even be hidden from the client application by interposing a
proxy server that performs this operation on behalf of the client.
<p>

Today, Venti provides little security.  After authenticating to the
server, clients can read any block for which they know the
fingerprint.  A fingerprint does act as a capability since the space
of fingerprints is large and the Venti protocol does not include a
means of enumerating the blocks on the server.  However, this
protection is weak as a single root fingerprint enables access to an
entire file tree and once a fingerprint is known, there is no way to
restrict access to a particular user.  We are exploring ways of
providing better access control.
<p>

To date, the structures we have used for storing data on Venti break
files into a series of fixed sized blocks.  Identical blocks are
consolidated on Venti, but this consolidation will not occur if the
data is shifted within the file or an application uses a different
block size.  This limitation can be overcome using an adaptation of
Manber's algorithm for finding similarities in files [9].  The idea is
to break files into variable sized blocks based on the identification
of anchor or break points, increasing the occurrence of duplicate
blocks [12].  Such a strategy can be implemented in client
applications with no change to the Venti server.
<p>

A more detailed analysis of the decade of daily snapshots of the Plan
9 file systems might be interesting.  The trace data we have made
publicly available contains approximately the same information used
for other studies of long term file activity [4].
<p>

<h1>10.  Conclusion</h1>
<p>

The approach of identifying a block by the Sha1 hash of its contents
is well suited to archival storage.  The write-once model and the
ability to coalesce duplicate copies of a block makes Venti a useful
building block for a number of interesting storage applications.
<p>

The large capacity of magnetic disks allows archival data to be
retained and available on-line with performance that is comparable to
conventional disks.  Stored on our prototype server is over a decade
of daily snapshots of two major departmental file servers.  These
snapshots are stored in a little over 200 Gbytes of disk space.
Today, 100 Gbytes drives cost less than $300 and IDE RAID controllers
are included on many motherboards.  A scaled down version of our
server could provide archival storage for a home user at an attractive
price.  Tomorrow, when terabyte disks can be had for the same price,
it seems unlikely that archival data will be deleted to reclaim space.
Venti provides an attractive approach to storing that data.
<p>

<h1>11.  Acknowledgments</h1>
<p>

This paper was improved by comments and suggestions from Peter Bosch,
Eric Grosse, Lorenz Huelsbergen, Rob Pike, Ross Quinlan, and Cliff
Young and six anonymous reviewers.  The paper's shepherd was Ethan L.
Miller.  We thank them all for their help.
<p>

<h1>12.  References</h1>
<p>

[1] Ann Chervenak, Vivekenand Vellanki, and Zachary Kurmas.
Protecting file systems: A survey of backup techniques.  In
Proceedings Joint NASA and IEEE Mass Storage Conference, March 1998.
<p>

[2] Arturo Crespo and Hector Garcia-Molina.  Archival storage for
digital libraries.  In Proceedings of the 3rd ACM International
Conference on Digital Libraries, 1998.
<p>

[3] Kevin Fu, Frans Kaashoek, and David Mazières.  Fast and secure
distributed read-only file system.  In Proceedings of the 4th
Symposium on Operating Systems Design and Implementation, 2000.
<p>

[4] Timothy J. Gibson, Ethan L. Miller, and Darrell D. E. Long.
Long-term file activity and inter-reference patterns.  In Proceedings,
24th International Conference on Technology Management and Performance
Evaluation of Enterprise-Wide Information Systems, Computer
Measurement Group, December 1998.
<p>

[5] Dave Hitz, James Lau, and Michael Malcolm, File system design for
an NFS file server appliance, In Proceedings of the Winter 1994 USENIX
Conference, San Francisco, CA, January 1994.
<p>

[6] J. K. Hollingsworth and E. L. Miller.  Using content-derived names
for configuration management.  In Proceeding of the 1997 ACM Symposium
on Software Reusability, Boston, May 1997.
<p>

[7] John Howard, Michael Kazar, Sherri Menees, David Nichols, Mahadev
Satyanarayanan, Robert Sidebotham, and Michael West.  Scale and
performance in a distributed file system.  ACM Transactions on
Computer Systems, 6(1):51-81, February 1988.
<p>

[8] Norman C. Hutchinson, Stephen Manley, Mike Federwisch, Guy Harris,
Dave Hitz, Steven Kleiman, and Sean O'Malley.  Logical vs.  physical
file system backup.  In Proceedings of the 3rd USENIX Symposium on
Operating Systems Design and Implementation (OSDI), 1999.
<p>

[9] Udi Manber.  Finding similar files in a large file system.  In
Proceedings of the Winter 1994 USENIX Conference, San Francisco, CA,
January 1994.
<p>

[10] Alfred J. Menezes, Paul C. van Oorschot, and Scott A. Vanstone.
Handbook of Applied Cryptography.  CRC Press, 1996.
<p>

[11] Ralph C. Merkle.  Protocols for public-key cryptosystems.  In
Proceedings of the IEEE Symposium on Security and Privacy, pp.
122-133, April 1980.
<p>

[12] Athicha Muthitacharoen, Benjie Chen, and David Mazières.  A
low-bandwidth network file system.  In Proceedings of the 18th
Symposium on Operating Systems Principles, October 2001.
<p>

[13] National Institute of Standards and Technology, FIPS 180-1.
Secure Hash Standard.  US Department of Commerce, April 1995.
<p>

[14] National Institute of Standards and Technology, Draft FIPS 180-2.
Secure Hash Standard.  US Department of Commerce, May 2001.
<p>

[15] Evi Nemeth, Garth Snyder, Scott Seebass, and Trent R. Hein.  UNIX
System Administration Handbook 3rd Edition, Prentice Hall, 2001.
<p>

[16] Rob Pike, Dave Presotto, Sean Dorward, Bob Flandrena, Ken
Thompson, Howard Trickey, and Phil Winterbottom.  Plan 9 from Bell
Labs, Computing Systems, Vol. 8, 3, pp.  221-254, Summer 1995.
<p>

[17] Sean Quinlan.  A cache worm file system.  Software-Practice and
Experience, Vol 21, 12, pp 1289-1299, December 1991.
<p>

[18] Douglas S. Santry, Michael J. Feeley, Norman C. Hutchinson,
Alistair C. Veitch, Ross W. Carton and Jacob Ofir.  Deciding when to
forget in the Elephant file system.  In Proceedings of the 17th
Symposium on Operating Systems Principles, December 12-15, 1999.
<p>

[19] John.  D. Strunk, Garth R. Goodson, Michael L. Scheinholtz, Craig
A.N. Soules, and Gregory R. Ganger.  Self-securing storage: protecting
data in compromised systems.  In Proceedings of the 4th Symposium on
Operating Systems Design and Implementation, October 2000.
<p>

[20] D. A. Thompson and J. S. Best.  The future of magnetic data
storage technology, IBM Journal of Research and Development, Vol 44,
3, pp.  311-322, May 2000.
<p>

[21] J. Ziv and A. Lempel.  A universal algorithm for sequential data
compression, IEEE Trans.  Inform.  Theory, vol.  IT-23, pp.  337-343,
May 1977.
<p>