To BLOB or Not To BLOB:
Large Object Storage in a Database or a Filesystem?
Russell Sears2, Catharine van Ingen1, Jim Gray1
1: Microsoft Research, 2: University of California at Berkeley
firstname.lastname@example.org, vanIngen@microsoft.com, email@example.com
Revised June 2006
One Microsoft Way
Redmond, WA 98052
To BLOB or Not To BLOB:
Large Object Storage in a Database or a Filesystem?
Russell Sears2, Catharine van Ingen1, Jim Gray1
1: Microsoft Research, 2: University of California at Berkeley
firstname.lastname@example.org, vanIngen@microsoft.com, email@example.com
April 2006 Revised June 2006
Application designers must decide whether to store
large objects (BLOBs) in a filesystem or in a database.
Generally, this decision is based on factors such as
application simplicity or manageability. Often, system
performance affects these factors.
Folklore tells us that databases efficiently handle
large numbers of small objects, while filesystems are
more efficient for large objects. Where is the
break-even point? When is accessing a BLOB stored
as a file cheaper than accessing a BLOB stored as a
Of course, this depends on the particular
filesystem, database system, and workload in question.
This study shows that when comparing the NTFS file
system and SQL Server 2005 database system on a
workload, BLOBs smaller than 256KB are more
efficiently handled by SQL Server, while NTFS is
more efficient BLOBS larger than 1MB. Of course,
this break-even point will vary among different
database systems, filesystems, and workloads.
By measuring the performance of a storage server
workload typical of web applications which use get/put
protocols such as WebDAV [WebDAV], we found that
the break-even point depends on many factors.
However, our experiments suggest that storage age, the
ratio of bytes in deleted or replaced objects to bytes in
live objects, is dominant. As storage age increases,
fragmentation tends to increase. The filesystem we
study has better fragmentation control than the
database we used, suggesting the database system
would benefit from incorporating ideas from filesystem
architecture. Conversely, filesystem performance may
be improved by using database techniques to handle
Surprisingly, for these studies, when average
object size is held constant, the distribution of object
sizes did not significantly affect performance. We also
found that, in addition to low percentage free space, a
low ratio of free space to average object size leads to
fragmentation and performance degradation.
Application data objects are getting larger as digital
media becomes ubiquitous. Furthermore, the
increasing popularity of web services and other
network applications means that systems that once
managed static archives of “finished” objects now
manage frequently modified versions of application
data as it is being created and updated. Rather than
updating these objects, the archive either stores
multiple versions of the objects (the V of WebDAV
stands for “versioning”), or simply does wholesale
replacement (as in SharePoint Team Services
Application designers have the choice of storing
large objects as files in the filesystem, as BLOBs
(binary large objects) in a database, or as a
combination of both. Only folklore is available
regarding the tradeoffs – often the design decision is
based on which technology the designer knows best.
Most designers will tell you that a database is probably
best for small binary objects and that that files are best
for large objects. But, what is the break-even point?
What are the tradeoffs?
This article characterizes the performance of an
abstracted write-intensive web application that deals
with relatively large objects. Two versions of the
system are compared; one uses a relational database to
store large objects, while the other version stores the
objects as files in the filesystem. We measure how
performance changes over time as the storage becomes
fragmented. The article concludes by describing and
quantifying the factors that a designer should consider
when picking a storage system. It also suggests
filesystem and database improvements for large object
One surprising (to us at least) conclusion of our
work is that storage fragmentation is the main
determinant of the break-even point in the tradeoff.
Therefore, much of our work and much of this article
focuses on storage fragmentation issues. In essence,
filesystems seem to have better fragmentation handling
than databases and this drives the break-even point
down from about 1MB to about 256KB.
Filesystems have long used sophisticated allocation
strategies to avoid fragmentation while laying out
objects on disk. For example, OS/360’s filesystem was
extent based and clustered extents to improve access
time. The VMS filesystem included similar
optimizations and provided a file attribute that allowed
users to request a (best effort) contiguous layout
[McCoy, Goldstein]. Berkeley FFS [McKusick] was an
early UNIX filesystem that took sequential access, seek
performance and other hardware characteristics into
account when laying data out on disk. Subsequent
filesystems were built with similar goals in mind.
The filesystem used in these experiments, NTFS,
uses a ‘banded’ allocation strategy for metadata, but
not for file contents [NTFS]. NTFS allocates space for
file stream data from a run-based lookup cache. Runs
of contiguous free clusters are ordered in decreasing
size and volume offset. NTFS attempts to satisfy a new
space allocation from the outer band. If that fails, large
extents within the free space cache are used. If that
fails, the file is fragmented. Additionally, when a file is
deleted, the allocated space cannot be reused
immediately; the NTFS transactional log entry must be
committed before the freed space can be reallocated.
The net behavior is that file stream data tends to be
allocated contiguously within a file.
In contrast, database systems began to support
large objects more recently [BLOB, DeWitt].
Databases historically focused on small (100 byte)
records and on clustering tuples within the same table.
Clustered indexes let users control the order in which
tuples are stored, allowing common queries to be
serviced with a sequential scan over the data.
Filesystems and databases
approaches to modifying
Filesystems are optimized for appending or truncating
a file. In-place file updates are efficient, but when data
are inserted or deleted in the middle of a file, all
contents after the modification must be completely
rewritten. Some databases completely rewrite modified
BLOBS; this rewrite is transparent to the application.
Others, such as SQL Server, adopt the Exodus design
that supports efficient insertion or deletion within an
object by using B-Tree based storage of large objects
IBM DB2’s DataLinksTM technology stores
BLOBs in the filesystem, and uses the database as an
associative index of the file metadata [Bhattacharya].
Their files are updated atomically using a mechanism
similar to old-master, new-master safe writes (Section
To ameliorate the fact that the database poorly
handles large fragmented objects, the application could
do its own de-fragmentation or garbage collection.
Some applications simply store each object directly in
the database as a single BLOB, or as a single file in a
filesystem. However, in order to efficiently service
user requests, other applications use more complex
allocation strategies. For instance, objects may be
partitioned into many smaller chunks stored as
BLOBS. Video streams are often “chunked” in this
way. Alternately, smaller objects may be aggregated
into a single BLOB or file – for example TAR, ZIP, or
2.2. Safe writes
This study only considers applications that insert,
replace or delete entire objects. In the replacement
case, the application creates a new version and then
deletes the original. Such applications do not force the
storage system to understand their data’s structure,
trading opportunities for optimization for simplicity
and robustness. Even this simple update policy is not
entirely straightforward. This section describes
mechanisms that safely update entire objects at once.
Most filesystems protect internal metadata
structures (such as directories and filenames) from
corruption due to dirty shutdowns (such as system
crashes and power outages). However, there is no such
guarantee for file contents. In particular, filesystems
and the operating system below them reorder write
requests to improve performance. Only some of those
requests may complete at dirty shutdown.
As a result, many desktop applications use a
technique called a “safe write” to achieve the property
that, after a dirty shutdown, the file contents are either
new or old, but not a mix of old and new. While
safe-writes ensure that an old file is robustly replaced,
they also force a complete copy of the file to disk even
if most of the file contents are unchanged.
To perform a safe write, a new version of the file
is created with a temporary name. Next, the new data
is written to the temporary file and those writes are
flushed to disk. Finally, the new file is renamed to the
permanent file name, thereby deleting the file with the
older data. Under UNIX, rename() is guaranteed to
atomically overwrite the old version of the file. Under
Windows, the ReplaceFile() call is used to atomically
replace one file with another.
In contrast, databases typically offer transactional
updates, which allow applications to group many
changes into a single atomic unit called a transaction.
Complete transactions are applied to the database
atomically regardless of system dirty shutdown,
database crash, or application crashes. Therefore,
applications may safely update their data in whatever
manner is most convenient. Depending on the database
implementation and the nature of the update, it can be
more efficient to update a small portion of the BLOB
instead of overwriting it completely.
The database guarantees transactional semantics
by logging both metadata and data changes throughout
a transaction. Once the appropriate log records have
been flushed to disk, the associated changes to the
database are guaranteed to complete—the write-ahead-
log protocol. The log is written sequentially, and
subsequent database disk writes can be reordered to
minimize seeks. Log entries for each change must be
written; but the database writes can be coalesced.
Logging all data and metadata guarantees
correctness at the cost of writing all data twice—once
to the database and once to the log. With large data
objects, this can put the database at a noticeable
performance disadvantage. The sequential write to the
database log is roughly equivalent to the sequential
write to a file. The additional write to the database
pages will form an additional sequential write if all
database pages are contiguous or many seek-intensive
writes if the pages are not located near each other.
Large objects also fill the log causing the need for
more frequent log truncation and checkpoint, which
reduces opportunities to reorder and combine page
We exploited the bulk-logged option of SQL
Server to ameliorate this penalty. Bulk-logged is a
special case for BLOB allocation and de-allocation – it
does not log the initial writes to BLOBs and their final
deletes; rather it just logs the meta-data changes (the
allocation or de-allocation information). Then when
the transaction commits the BLOB changes are
committed. This supports transaction UNDO and
supports warm restart, but it does not support media
recovery (there is no second-copy of the BLOB data in
the log) – it is little-D ACId rather than big-D ACID.
2.3. Dealing with Media Failures
Databases and filesystems both provide transactional
updates to object metadata. When the bulk-logged SQL
option is used, both kinds of systems write the large
object data once. However, database systems guarantee
that large object data are updated atomically across
system or application outages. However, neither
protects against media failures.
Media failure detection, protection, and repair
mechanisms are often built into large-scale web
services. Such services are typically supported by very
large, heterogeneous networks built from low cost,
application errors that silently corrupt data are not
uncommon. Applications deployed on such systems
must regularly check for and recover from corrupted
application data. When corruption is detected, intact
Hardware faults and
versions of the affected data can be copied from a valid
This study does not consider the overhead of
detecting and correcting silent data corruption and
media failures. These overheads are comparable for
the both file and database systems.
3. Prior work
While filesystem fragmentation has been studied
within numerous contexts in the past, we were
surprised to find relatively
investigations. There is common folklore passed by
word of mouth between application and system
designers. A number of performance benchmarks are
impacted by fragmentation, but these do not measure
fragmentation per se. There are algorithms used for
space allocation by database and filesystem designers.
We found very little hard data on the actual
performance of fragmented storage. Moreover, recent
changes in application workloads, hardware models,
and the increasing popularity of database systems for
the storage of large objects present workloads not
covered by existing studies.
There is a wealth of anecdotal experience with
applications that use large objects. The prevailing
wisdom is that databases are better for small objects
while filesystems are better for large objects. The
boundary between small and large is usually a bit
fuzzy. The usual reasoning is:
Database queries are faster than file opens. The
overhead of opening a file handle dominates
performance when dealing with small objects.
Reading or writing large files is faster than
accessing large database BLOBS. Filesystems are
optimized for streaming large objects.
Database client interfaces aren’t good with large
objects. Remote database client interfaces such as
MDAC have historically been optimized for short
low latency requests returning small amounts of
File opens are CPU expensive, but can be easily
amortized over cost of streaming large objects.
None of the above points address the question of
application complexity. Applications that store large
objects in the filesystem encounter the question of how
to keep the database object metadata and the filesystem
object data synchronized. A common problem is the
garbage collection of files that have been “deleted” in
the database but not the filesystem.
Also missing are operational issues such as
replication, backup, disaster
3.2. Standard Benchmarks
While many filesystem benchmarking tools exist, most
consider the performance of clean filesystems, and do
not evaluate long-term performance as the storage
system ages and fragments. Using simple clean initial
conditions eliminates potential variation in results
caused by different initial conditions and reduces the
need for settling time to allow the system to reach
Several long-term filesystem performance studies
have been performed based upon two general
approaches [Seltzer]. The first approach, trace based
load generation, uses data gathered from production
systems over a long period. The second approach, and
the one we adopt for our study, is vector based load
generation that models application behavior as a list of
primitives (the ‘vector’), and randomly applies each
primitive to the filesystem with the frequency a real
application would apply the primitive to the system.
NetBench [NetBench] is the most common
Windows file server benchmark. It measures the
performance of a file server accessed by multiple
clients. NetBench generates workloads typical of
SPC-2 benchmarks storage system applications
that read and write large files in place, execute large
read-only database queries, or provide read-only
on-demand access to video files [SPC].
The Transaction Processing Performance Council
[TPC] have defined several benchmark suites to
characterize online transaction processing workloads
and also decision support workloads. However, these
benchmarks do not capture the task of managing large
objects or multimedia databases.
None of these benchmarks
3.3. Data layout mechanisms
approaches to the fragmentation problem.
The creators of FFS observed that for typical
workloads of the time, fragmentation avoiding
allocation algorithms kept fragmentation under control
as long as volumes were kept under 90% full [Smith].
UNIX variants still reserve a certain amount of free
space on the drive, both for disaster recovery and in
order to prevent excess fragmentation.
NTFS disk occupancy on deployed Windows
systems varies widely. System administrators’ target
disk occupancy may be as low as 60% or over 90%
[NTFS]. On NT 4.0, the in-box defragmentation utility
was known to have difficulties running when the
occupancy was greater than 75%. This limitation was
addressed in subsequent releases. By Windows 2003
SP1, the utility included support for defragmentation of
systems take surprisingly different
system files and attempted partial file defragmentation
when full defragmentation is not possible.
LFS [Rosenblum], a log based filesystem,
optimizes for write performance by organizing data on
disk according to the chronological order of the write
requests. This allows it to service write requests
sequentially, but causes severe fragmentation when
files are updated randomly. A cleaner that
simultaneously defragments the disk and reclaims
deleted file space can partially address this problem.
Network Appliance’s WAFL (“Write Anywhere
File Layout”) [Hitz] is able to switch between
conventional and write-optimized
depending on workload conditions. WAFL also
leverages NVRAM caching for efficiency and provides
access to snapshots of older versions of the filesystem
contents. Rather than a direct copy-on-write of the
data, WAFL metadata remaps the file blocks. A
defragmentation utility is supported, but is said not to
be needed until disk occupancy exceeds 90+%.
GFS [Ghemawat], a filesystem designed to deal
with multi-gigabyte files on 100+ terabyte volumes,
partially addresses the data layout problem by using
64MB blocks called ‘chunks’. GFS also provides a
safe record append operation that allows many clients
to simultaneously append to the same file, reducing the
number of files (and opportunities for fragmentation)
exposed to the underlying filesystem. GFS records
may not span chunks, which can result in internal
fragmentation. If the application attempts to append a
record that will not fit into the end of the current
chunk, that chunk is zero padded, and the new record is
allocated at the beginning of a new chunk. Records are
constrained to be less than ¼ the chunk size to prevent
excessive internal fragmentation. However, GFS does
not explicitly attempt to address fragmentation
introduced by the underlying filesystem, or to reduce
internal fragmentation after records are allocated.
Comparing Files and
This study is primarily concerned with the deployment
and performance of data-intensive web services.
Therefore, we opted for a simple vector based
workload typical of existing web applications, such as
WebDAV applications, SharePoint, Hotmail, flickr,
and MSN spaces. These sites allow sharing of objects
that range from small text mail messages (100s of
bytes) to photographs (100s of KB to a few MB) to
video (100s of MBs.)
The workload also corresponds to collaboration
applications such as SharePoint Team Services
[SharePoint]. These applications enable rich document
sharing semantics, rather than simple file shares.
Examples of the extra semantics include document
versioning, content indexing and search, and rich
Many of these sites and applications use a database
to hold the application specific metadata including that
which describes the large objects. Large objects may be
stored in the database, in the filesystem, or distributed
between the database and the filesystem. We consider
only the first two options here. We also did not include
any shredding or chunking of the objects.
4.1. Test System Configuration
All the tests were performed on the system described in
Table 1: All test code was written using C# in Visual
Studio 2005 Beta 2. All binaries used to generate tests
were compiled to x86 code with debugging disabled.
Table 1: Configurations of the test systems
Tyan S2882 K8S Motherboard,
1.8 Ghz Opteron 244, 2 GB RAM (ECC)
SuperMicro “Marvell” MV8 SATA controller
4 Seagate 400GB ST3400832AS 7200 rpm SATA
Windows Server 2003 R2 Beta (32 bit mode).
SQL Server 2005 Beta 2 (32 bit mode).
4.2. File based storage
For the filesystem based storage tests, we stored
metadata such as object names and replica locations in
SQL server tables. Each application object was stored
in its own file. The files were placed in a single
directory on an otherwise empty NTFS volume. SQL
was given a dedicated log and data drive, and the
NTFS volume was accessed via an SMB share.
We considered a purely file-based approach, but
such a system would need to implement complex
recovery routines, would lack support for consistent
metadata, and would not provide functionality
comparable to the systems mentioned above.
This partitioning of tasks between a database and
filesystem is fairly flexible, and allows a number of
replication and load balancing schemes. For example, a
single clustered SQL server could be associated with
several file servers. Alternately, the SQL server could
be co-located with the associated files and then the
combination clustered. The database isolates the client
from changes in the architecture – changing the pointer
in the database changes the path returned to the client.
We chose to measure the configuration with the
database co-located with the associated files. This
single machine configuration kept our experiments
simple and avoids building
dependencies on the network layout into the study.
However, we structured all code to use the same
interfaces and services as a networked configuration.
4.3. Database storage
The database storage tests were designed to be as
similar to the filesystem tests as possible. As explained
previously, we used bulk-logging mode. We also used
out-of-row storage for the BLOB data so that the
BLOBS did not decluster the metadata.
For simplicity, we did not create the table with the
option TEXTIMAGE ON filegroup, which would
store BLOBS in a separate filegroup. Although the
BLOB data and table information are stored in the
same file group, out-of-row storage places BLOB data
on pages that are distinct from the pages that store the
other table fields. This allows the table data to be kept
in cache even if the BLOB data does not fit in main
memory. Analogous table schemas and indices were
used and only minimal changes were made to the
software that performed the tests.
4.4. Performance tuning
We set out to fairly evaluate the out-of-the-box
performance of the two storage systems. Therefore, we
did no performance tuning except in cases where the
default settings introduced gross discrepancies in the
functionality that the two systems provided.
We found that NTFS’s file allocation routines
behave differently when the number of bytes per file
write (append) is varied. We did not preset the file
size; as such NTFS attempts to allocate space as
needed. When NTFS detects large, sequential appends
to the end of a file its allocation routines aggressively
attempt to allocate contiguous space. Therefore, as the
disk gets full, smaller writes are likely to lead to more
fragmentation. While NTFS supports setting the valid
data length of a file, this operation incurs the write
overhead of zero filling the file, so it is not of interest
here [NTFS]. Because the behavior of the allocation
routines depends on the size of the write requests, we
use a 64K write buffer for all database and filesystem
runs. The files were accessed (read and written)
sequentially. We made no attempt to pass hints
regarding final object sizes to NTFS or SQL Server.
4.5. Workload generation
Real world workloads have many properties that are
difficult to model without application specific
information such as object size distributions, workload
characteristics, or application traces.
However, the applications of interest are extremely
simple from a storage replica’s point of view. Over
time, a series of object allocation, deletion, and
safe-write updates are processed with interleaved read
For simplicity, we assumed that all objects are
equally likely to be written and/or read. We also
assumed that there is no correlation between objects.
This lets us measure the performance of write-only and
We measured constant size objects rather than
objects with more complicated size distributions. We
expected that size distribution would be an important
factor in our experiments. As shown later, we found
that size distribution had no obvious effect on the
behavior. Given that, we chose the very simple
constant size distribution.
Before continuing, we should note that synthetic
load generation has led to misleading results in the
past. We believe that this is less of an issue with the
applications we study, as random storage partitioning
and other load balancing techniques remove much of
the structure from application workloads. However,
our results should be validated by comparing them to
results from a trace-based benchmark. At any rate,
fragmentation benchmarks that rely on traces are of
little use before an application has been deployed, so
synthetic workloads may be the best tool available to
web service designers.
4.6. Storage age
We want to characterize the behavior of storage
systems over time. Past fragmentation studies measure
age in elapsed time such as ‘days’ or ‘months.’ We
propose a new metric, storage age, which is the ratio of
bytes in objects that once existed on a volume to the
number of bytes in use on the volume. This definition
of storage age assumes that the amount of free space on
a volume is relatively constant over time. This
measurement is similar to measurements of heap age
for memory allocators, which frequently measure time
in “bytes allocated,” or “allocation events.”
For a safe-write system storage age is the ratio of
object insert-update-delete bytes over the number of
total object bytes. When evaluating a single node
system using trace-based data, storage age has a simple
A synthetic workload effectively speeds up
application elapsed time. Our workload is disk arm
limited; we did not include extra think time or
processor overheads. The speed up is application
specific and depends on the actual application
read:write ratio and heat.
We considered reporting time in “hours under
load”. Doing so has the undesirable property of
rewarding slow storage systems by allowing them to
perform less work during the test.
In our experimental setup, storage age is
equivalent to “safe writes per object.” This metric is
independent of the actual applied read:write load or the
number of requests over time. Storage ages can be
compared across hardware
applications. Finally, it is easy to convert storage age
to elapsed wall clock time after the rate of data churn,
or overwrite rate, of a specific system is determined.
Our results use throughput as the primary indicator of
performance. We started
out-of-the-box throughput study. We then looked at the
longer term changes caused by fragmentation with a
focus on 256K to 1M object sizes where the filesystem
and database have comparable performance. Lastly,
we discuss the effects of volume size and object size on
with the typical
5.1. Database or Filesystem: Throughput
We begin by establishing when a database is clearly the
right answer and when the filesystem is clearly the
Following the lead of existing benchmarks, we
evaluated the read performance of the two systems on a
clean data store. Figure 1 demonstrates the truth of the
folklore: objects up to about 1MB are best stored as
database BLOBS. Performance of SQL reads for
objects of 256KB was 2x better than NTFS; but the
systems had parity at 1MB objects. With 10MB
objects, NTFS outperformed SQL Server.
The write throughput of SQL Server exceeded that
of NTFS during bulk load. With 512KB objects,
database write throughput was 17.7MB/s, while the
filesystem only achieved 10.1MB/s.
5.2 Database or Filesystem over time
Next, we evaluated the performance of the two systems
on large objects over time. If fragmentation is
important, we expect to see noticeable performance
Our first discovery was that SQL Server does not
Read Throughput After Bulk load
Figure 1: Read throughput immediately after bulk
loading the data. Databases are fastest on small
objects. As object size increases, NTFS throughput
improves faster than SQL Server throughput.
provide facilities to report fragmentation of large object
data or to defragment such data. While there are
measurements and mechanisms
defragmentation, the recommended way to defragment
a large BLOB table is to create a new table in a new
file group, copy the old records to the new table and
drop the old table [SQL].
To measure fragmentation, we tagged each of our
objects with a unique identifier and a sequence number
at 1KB intervals. We also implemented a utility that
looks for the locations of these markers on a raw
device in a way that is robust to page headers and other
artifacts of the storage system. In other words, the
utility measures fragmentation in the same way
regardless of whether the objects are stored in the
filesystem or the database. We ran our utility against
an NTFS volume to check that it reported figures that
agreed with the NTFS fragmentation report.
The degradation in read performance for 256K,
512K, and 1MB BLOBS is shown in Figure 2. Each
storage age (2 and 4) corresponds to the time necessary
for the number of updates, inserts, or deletes to be N
times the number of objects in our store since the bulk
load (storage age 0) shown in Figure 1. Fragmentation
under NTFS begins to level off over time. SQL
Server’s fragmentation increases almost linearly over
time and does not seem to be approaching any
asymptote. This is dramatically the case with very
large (10MB) objects as seen in Figure 3.
We also ran our load generator against an
artificially and pathologically fragmented NTFS
volume. We found that fragmentation slowly decreased
over time. The best-effort attempt to allocate
contiguous space actually defragments such volumes.
This suggests that NTFS is indeed approaching an
asymptote in Figure 3.
Figure 4 shows the degradation in write
performance as storage age increases. In both database
and filesystems, the write throughput during bulk load
is much better than read throughput immediately
afterward. This is not surprising, as the storage
systems can simply append each new file to the end of
allocated storage, avoiding seeks during bulk load. On
the other hand, the read requests are randomized, incur
the overhead of at least one seek. After bulk load, the
write performance of SQL Server degrades quickly,
while the NTFS write performance numbers are
slightly better than its read performance.
Read Throughput After Two Overwrites
Read Throughput After Four Overwrites
Figure 2: Fragmentation causes read performance
to degrade over time. The filesystem is less
affected by this than SQL Server. Over time
NTFS outperforms SQL Server when objects are
larger than 256KB.
Long Term Fragmentation With 10 MB Objects
Figure 3: For large objects, NTFS deals with
fragmentation more effectively than SQL server.
“Storage Age” is the average number of times
each object has been replaced with a newer
version. An object in a contiguous region on disk
has 1 fragment.
512K Write Throughput Over Time
After bulk load (zero)TwoFour
Figure 4: Although SQL Server quickly fills a
volume with data, performance suffers when
existing objects are replaced.
Note that these write performance numbers are not
directly comparable to the read performance numbers
in Figures 1 and 2. Read performance is measured after
fragmentation, while write performance is the average
performance during fragmentation. To be clear, the
“storage age four” write performance is the average
write throughput between the read measurements
labeled “bulk load” and “storage age two.” Similarly,
the reported write performance for storage age four
reflects average write performance between storage
ages two and four.
The results so far indicate that as storage age
increases, the parity point where filesystems and
databases have comparable performance declines from
1MB to 256KB. Objects up to about 256KB are best
kept in the database; larger objects should be in the
To verify this, we attempted to run both systems
until the performance reached a steady state.
Figure 5 indicates that fragmentation converges to
four fragments per file, or one fragment per 64KB, in
both the filesystem and database. This is interesting
because our tests use 64KB write requests, again
suggesting that the impact of write block size upon
fragmentation warrants further study. From this data,
we conclude that SQL Server indeed outperforms
NTFS on objects under 256KB, as indicated by
Figures 2 and 4.
5.3. Fragmentation effects of object size,
volume capacity, and write request size
Distributions of object size vary greatly from
application to application. Similarly, applications are
deployed on storage volumes of widely varying size
particularly as disk capacity continues to increase.
This series of tests generated objects using a
constant size distribution and compared performance
when the sizes were uniformly distributed. Both sets
of objects had a mean size of 10MB.
Intuition suggested that constant size objects
should not lead to fragmentation. Deleting an initially
contiguous object leaves a region of contiguous free
space exactly the right size for any new object. As
shown in Figure 6, our intuition was wrong.
As long as the average object size is held constant
there is little difference between uniformly distributed
and constant sized objects. This suggests that
experiments that use
distributions can be representative of many different
workloads. This contradicts the approach taken by
prior storage benchmarks that make use of complex,
accurate modeling of application workloads. This may
well be due to the simple all-or-nothing access pattern
that avoids object extension and truncation, and our
assumption that application code has not been carefully
tuned to match the underlying storage system.
The time it takes to run the experiments is
proportional to the volume’s capacity. When the entire
disk capacity (400GB) is used, some experiments take
a week to complete. Using a smaller (although perhaps
unrealistic) volume size, allows more experiments; but
how trustworthy are the results?
As shown in Figure 7, we found that volume size
does not really affect performance for larger volume
sizes. However, on smaller volumes, we found that as
the ratio of free space to object size decreases,
extremely simple size
Long Term Fragmentation With 256K Objects
Figure 5: For small objects, the systems have
similar fragmentation behavior.
Database Fragmentation: Blob Distributions
Filesystem Fragmentation: Blob Distributions
Figure 6: Fragmentation for large (10MB) BLOBS
– increases slowly for NTFS but rapidly for SQL.
However objects of a constant size show no better
fragmentation performance than objects of sizes
chosen uniformly at random with the same
We did not characterize the exact point where this
becomes a significant issue. However, our results
suggest that the effect is negligible when there is 10%
free space on a 40GB volume storing 10MB objects,
which implies a pool of 400 free objects. With a 4GB
volume with a pool of 40 free objects, performance
While these findings hold for the SQL Server and
NTFS allocation policies, they probably do not hold for
all current production systems. Characterizing other
allocation policies is beyond the scope of this work.
6. Implications for system
This article has already mentioned several issues that
should be considered during application design.
Designers should provision at least 10% excess storage
capacity to allow each volume to maintain free space
for many (~400 in our experiment) free objects. If the
volume is large enough, the percentage free space
becomes a limiting factor. For NTFS, we can see this
in Figure 7, where the performance of a 97.5% full
400GB volume is worse than the performance of a 90%
full 40GB volume. (A 99% full 400GB volume would
have the same number of free objects as the 40GB
While we did not carefully characterize the impact
of application allocation routines upon the allocation
strategy used by the underlying storage system, we did
observe significant differences in behavior as we varied
the write buffer size. Experimentation with different
buffer sizes or other techniques that avoid incremental
allocation of storage may significantly improve long
run storage performance. This also suggests that
filesystem designers should re-evaluate what is a
“large” request and be more aggressive about
coalescing larger sequential requests.
Simple procedures such as manipulating write
size, increasing the amount of free space, and
performing periodic defragmentation can improve the
performance of a system. When dealing with an
existing system, tuning these parameters may be
preferable to switching from database to filesystem
storage, or vice versa.
When designing a new system, it is important to
consider the behavior of a system over time instead of
looking only the performance of a clean system. If
fragmentation is a significant concern, the system must
be defragmented regularly. Defragmentation of a
filesystem implies significant read/write impacts or
application logic to garbage collect and reinstantiate a
volume. Defragmentation of a database requires
explicit application logic to copy existing BLOBS into
a new table. To avoid causing still more
fragmentation, that logic must be run only when ample
free space is available. A good database
defragmentation utility (or at least good automation of
the above logic including space estimation required)
would clearly help system administrators.
Using storage age to measure time aids in the
comparison of different designs. In this study we use
“safe-writes per object” as a measurement of storage
age. In other applications, appends per object or some
combination of create/append/deletes may be more
For the synthetic workload presented above, NFTS
based storage works well for objects larger than
256KB. A better database BLOB implementation
would change this. At a minimum, the database should
report fragmentation. An in-place defragmentation
Database Fragmentation: Different Volumes
0 0.51 1.52 2.53 3.54 4.55
50% full - 40G
50% full - 400G
Filesystem Fragmentation: Different Volumes
50% full - 40G
50% full - 400G
Filesystem Fragmentation: Different Volumes
90% full - 40G
90% full - 400G
97.5% full - 40G
97.5% full - 400G
Figure 7: Fragmentation for 40GB and 400GB
volumes. Other than the 50% full filesystem run,
volume size has little impact on fragmentation.
utility would be helpful. To support incremental object
modification rather than the full rewrite considered
here, a more flexible B-Tree based BLOB storage
algorithm that optimizes insertion and deletion of
arbitrary data ranges within objects would be
The results presented here predict the performance of a
class of storage workloads, and reveal a number of
previously unknown factors in the importance of
storage fragmentation. They describe a simple
methodology that can measure the performance of
other applications that perform a limited number of
storage create, read, update, write, and delete
The study indicates that if objects are larger than
one megabyte on average, NTFS has a clear advantage
over SQL Server. If the objects are under 256
kilobytes, the database has a clear advantage. Inside
this range, it depends on how write intensive the
workload is, and the storage age of a typical replica in
Instead of providing measurements in wall clock
time, we use storage age, which makes it easy to apply
results from synthetic workloads to real deployments.
We are amazed that so little information regarding
the performance of fragmented storage was available.
Future studies should explore how fragmentation
changes under load. We did not investigate the
behavior of NTFS or SQL Server when multiple writes
to multiple objects are interleaved. This may happen if
objects are slowly appended to over long periods of
time or in multithreaded systems that simultaneously
create many objects. We expect that fragmentation gets
worse due to the competition, but how much worse?
Finally, this study only considers a single
filesystem and database implementation. We look
forward to seeing similar studies of other systems.
Any meaningful comparison
technologies should take fragmentation into account.
We thank Eric Brewer for the idea behind our
fragmentation analysis tool, helping us write this article
and reviewing several earlier drafts. We also thank
Surendra Verma, Michael Zwilling and the SQL Server
and NTFS development teams for answering numerous
questions throughout the study. We thank Dave
DeWitt, Ajay Kalhan, Norbert Kusters, Wei Xiao and
Michael Zwilling for their constructive criticism of the
[Bhattacharya] Suparna Bhattacharya, Karen W.
Brannon, Hui-I Hsiao, C. Mohan, Inderpal Narang
and Mahadevan Subramanian. “Coordinating
Backup/Recovery and Data Consistency Between
Database and File Systems. ACM SIGMOD, June 4-
[DeWitt] D. DeWitt, M. Carey, J. Richardson, E.
Shekita. “Object and File Management in the
EXODUS Extensible Database System.” VLDB,
Japan, August 1986.
[Ghemawat] S. Ghemawat, H. Goblioff, and S. Leung.
“The Google File System.” SOSP, October 19-21,
[Goldstein] A. Goldstein. “The Design and
Implementation of a Distributed File System.”
Digital Technical Journal, Number 5, September
[Hitz] D. Hitz, J. Lau and M. Malcom. “File System
Design for an NFS File Server Appliance.” NetApp
Technical Report #3002, March, 1995.
[McCoy] K. McCoy. VMS File System Internals.
Digital Press, 1990.
[McKusick] M. K. McKusick, W. N. Joy, S. J. Leffler,
Robert S. Fabry. “A Fast File System for UNIX.”
Computer Systems, Vol 2 #3 pages 181-197, 1984.
[NetBench] “NetBench.” Lionbridge Technologies,
[NTFS] Microsoft NTFS Development Team. Personal
communication. August, 2005.
[Rosenblum] M. Rosenblum, J. K. Ousterhout. “The
Design and Implementation of a Log-Structured File
System.” ACM TOCS, V. 10.1, pages 26-52, 1992.
[Seltzer] M. Seltzer, D. Krinsky, K. Smith, X. Zhang.
“The Case for Application-Specific Benchmarking.”
HotOS, page 102, 1999.
[Smith] K. Smith, M. Seltzer. “A Comparison of FFS
Disk Allocation Policies.” USENIX Annual
Technical Conference, 1996.
[SPC] SPC Benchmark-2 (SPC-2) Official
Specification, Version 1.0. Storage Performance
[SQL] Microsoft SQL Server Development Team.
Personal communication. August, 2005.
[TPC] Transaction Processing Performance Council.