The Design and Implementation of a Log-Structured File System
Mendel Rosenblum and John K. Ousterhout
Electrical Engineering and Computer Sciences, Computer Science Division
University of California
Berkeley, CA 94720
mendel@sprite.berkeley.edu, ouster@sprite.berkeley.edu

Abstract

magnitude more efficiently than current file systems.

This paper presents a new technique for disk storage
management called a log-structured file system. A logstructured file system writes all modifications to disk
sequentially in a log-like structure, thereby speeding up
both file writing and crash recovery. The log is the only
structure on disk; it contains indexing information so that
files can be read back from the log efficiently. In order to
maintain large free areas on disk for fast writing, we divide
the log into segments and use a segment cleaner to
compress the live information from heavily fragmented
segments. We present a series of simulations that demonstrate the efficiency of a simple cleaning policy based on
cost and benefit. We have implemented a prototype logstructured file system called Sprite LFS; it outperforms
current Unix file systems by an order of magnitude for
small-file writes while matching or exceeding Unix performance for reads and large writes. Even when the overhead
for cleaning is included, Sprite LFS can use 70% of the
disk bandwidth for writing, whereas Unix file systems typically can use only 5-10%.

Log-structured file systems are based on the assumption that files are cached in main memory and that increasing memory sizes will make the caches more and more
effective at satisfying read requests[1]. As a result, disk
traffic will become dominated by writes. A log-structured
file system writes all new information to disk in a sequential structure called the log. This approach increases write
performance dramatically by eliminating almost all seeks.
The sequential nature of the log also permits much faster
crash recovery: current Unix file systems typically must
scan the entire disk to restore consistency after a crash, but
a log-structured file system need only examine the most
recent portion of the log.
The notion of logging is not new, and a number of
recent file systems have incorporated a log as an auxiliary
structure to speed up writes and crash recovery[2, 3]. However, these other systems use the log only for temporary
storage; the permanent home for information is in a traditional random-access storage structure on disk. In contrast,
a log-structured file system stores data permanently in the
log: there is no other structure on disk. The log contains
indexing information so that files can be read back with
efficiency comparable to current file systems.

1. Introduction
Over the last decade CPU speeds have increased
dramatically while disk access times have only improved
slowly. This trend is likely to continue in the future and it
will cause more and more applications to become diskbound. To lessen the impact of this problem, we have devised a new disk storage management technique called a
log-structured file system, which uses disks an order of

For a log-structured file system to operate efficiently,
it must ensure that there are always large extents of free
space available for writing new data. This is the most
difficult challenge in the design of a log-structured file system. In this paper we present a solution based on large
extents called segments, where a segment cleaner process
continually regenerates empty segments by compressing
the live data from heavily fragmented segments. We used
a simulator to explore different cleaning policies and
discovered a simple but effective algorithm based on cost
and benefit: it segregates older, more slowly changing data
from young rapidly-changing data and treats them differently during cleaning.

The work described here was supported in part by the National Science Foundation under grant CCR-8900029, and in part
by the National Aeronautics and Space Administration and the
Defense Advanced Research Projects Agency under contract
NAG2-591.
This paper will appear in the Proceedings of the 13th ACM Symposium on Operating Systems Principles and the February 1992
ACM Transactions on Computer Systems.

July 24, 1991

We have constructed a prototype log-structured file
system called Sprite LFS, which is now in production use
as part of the Sprite network operating system[4]. Benchmark programs demonstrate that the raw writing speed of
Sprite LFS is more than an order of magnitude greater than
that of Unix for small files. Even for other workloads, such
-1-

as those including reads and large-file accesses, Sprite LFS
is at least as fast as Unix in all cases but one (files read
sequentially after being written randomly). We also measured the long-term overhead for cleaning in the production
system. Overall, Sprite LFS permits about 65-75% of a
disk’s raw bandwidth to be used for writing new data (the
rest is used for cleaning). For comparison, Unix systems
can only utilize 5-10% of a disk’s raw bandwidth for writing new data; the rest of the time is spent seeking.

and larger main memories make larger file caches possible.
This has two effects on file system behavior. First, larger
file caches alter the workload presented to the disk by
absorbing a greater fraction of the read requests[1, 6].
Most write requests must eventually be reflected on disk for
safety, so disk traffic (and disk performance) will become
more and more dominated by writes.

The remainder of this paper is organized into six sections. Section 2 reviews the issues in designing file systems for computers of the 1990’s. Section 3 discusses the
design alternatives for a log-structured file system and
derives the structure of Sprite LFS, with particular focus on
the cleaning mechanism. Section 4 describes the crash
recovery system for Sprite LFS. Section 5 evaluates Sprite
LFS using benchmark programs and long-term measurements of cleaning overhead. Section 6 compares Sprite
LFS to other file systems, and Section 7 concludes.

The second impact of large file caches is that they
can serve as write buffers where large numbers of modified
blocks can be collected before writing any of them to disk.
Buffering may make it possible to write the blocks more
efficiently, for example by writing them all in a single
sequential transfer with only one seek. Of course, writebuffering has the disadvantage of increasing the amount of
data lost during a crash. For this paper we will assume that
crashes are infrequent and that it is acceptable to lose a few
seconds or minutes of work in each crash; for applications
that require better crash recovery, non-volatile RAM may
be used for the write buffer.

2. Design for file systems of the 1990’s

2.2. Workloads

File system design is governed by two general
forces: technology, which provides a set of basic building
blocks, and workload, which determines a set of operations
that must be carried out efficiently. This section summarizes technology changes that are underway and describes
their impact on file system design. It also describes the
workloads that influenced the design of Sprite LFS and
shows how current file systems are ill-equipped to deal
with the workloads and technology changes.

Several different file system workloads are common
in computer applications. One of the most difficult workloads for file system designs to handle efficiently is found
in office and engineering environments. Office and
engineering applications tend to be dominated by accesses
to small files; several studies have measured mean file
sizes of only a few kilobytes[1, 6-8]. Small files usually
result in small random disk I/Os, and the creation and deletion times for such files are often dominated by updates to
file system ‘‘metadata’’ (the data structures used to locate
the attributes and blocks of the file).

2.1. Technology
Three components of technology are particularly
significant for file system design: processors, disks, and
main memory. Processors are significant because their
speed is increasing at a nearly exponential rate, and the
improvements seem likely to continue through much of the
1990’s. This puts pressure on all the other elements of the
computer system to speed up as well, so that the system
doesn’t become unbalanced.
Disk technology is also improving rapidly, but the
improvements have been primarily in the areas of cost and
capacity rather than performance. There are two components of disk performance: transfer bandwidth and
access time. Although both of these factors are improving,
the rate of improvement is much slower than for CPU
speed. Disk transfer bandwidth can be improved substantially with the use of disk arrays and parallel-head disks[5]
but no major improvements seem likely for access time (it
is determined by mechanical motions that are hard to
improve). If an application causes a sequence of small disk
transfers separated by seeks, then the application is not
likely to experience much speedup over the next ten years,
even with faster processors.
The third component of technology is main memory,
which is increasing in size at an exponential rate. Modern
file systems cache recently-used file data in main memory,
July 24, 1991
-2-

Workloads dominated by sequential accesses to large
files, such as those found in supercomputing environments,
also pose interesting problems, but not for file system
software. A number of techniques exist for ensuring that
such files are laid out sequentially on disk, so I/O performance tends to be limited by the bandwidth of the I/O and
memory subsystems rather than the file allocation policies.
In designing a log-structured file system we decided to
focus on the efficiency of small-file accesses, and leave it
to hardware designers to improve bandwidth for large-file
accesses. Fortunately, the techniques used in Sprite LFS
work well for large files as well as small ones.

2.3. Problems with existing file systems
Current file systems suffer from two general problems that make it hard for them to cope with the technologies and workloads of the 1990’s. First, they spread information around the disk in a way that causes too many small
accesses. For example, the Berkeley Unix fast file system
(Unix FFS)[9] is quite effective at laying out each file
sequentially on disk, but it physically separates different
files. Furthermore, the attributes (‘‘inode’’) for a file are
separate from the file’s contents, as is the directory entry
containing the file’s name. It takes at least five separate
disk I/Os, each preceded by a seek, to create a new file in

Unix FFS: two different accesses to the file’s attributes
plus one access each for the file’s data, the directory’s data,
and the directory’s attributes. When writing small files in
such a system, less than 5% of the disk’s potential
bandwidth is used for new data; the rest of the time is
spent seeking.

directories, and almost all the other information used to
manage the file system. For workloads that contain many
small files, a log-structured file system converts the many
small synchronous random writes of traditional file systems
into large asynchronous sequential transfers that can utilize
nearly 100% of the raw disk bandwidth.

The second problem with current file systems is that
they tend to write synchronously: the application must wait
for the write to complete, rather than continuing while the
write is handled in the background. For example even
though Unix FFS writes file data blocks asynchronously,
file system metadata structures such as directories and
inodes are written synchronously. For workloads with
many small files, the disk traffic is dominated by the synchronous metadata writes. Synchronous writes couple the
application’s performance to that of the disk and make it
hard for the application to benefit from faster CPUs. They
also defeat the potential use of the file cache as a write
buffer. Unfortunately, network file systems like NFS[10]
have introduced additional synchronous behavior where it
didn’t used to exist. This has simplified crash recovery, but
it has reduced write performance.

Although the basic idea of a log-structured file system is simple, there are two key issues that must be
resolved to achieve the potential benefits of the logging
approach. The first issue is how to retrieve information
from the log; this is the subject of Section 3.1 below. The
second issue is how to manage the free space on disk so
that large extents of free space are always available for
writing new data. This is a much more difficult issue; it is
the topic of Sections 3.2-3.6. Table 1 contains a summary
of the on-disk data structures used by Sprite LFS to solve
the above problems; the data structures are discussed in
detail in later sections of the paper.

3.1. File location and reading
Although the term ‘‘log-structured’’ might suggest
that sequential scans are required to retrieve information
from the log, this is not the case in Sprite LFS. Our goal
was to match or exceed the read performance of Unix FFS.
To accomplish this goal, Sprite LFS outputs index structures in the log to permit random-access retrievals. The
basic structures used by Sprite LFS are identical to those
used in Unix FFS: for each file there exists a data structure
called an inode, which contains the file’s attributes (type,
owner, permissions, etc.) plus the disk addresses of the
first ten blocks of the file; for files larger than ten blocks,
the inode also contains the disk addresses of one or more
indirect blocks, each of which contains the addresses of
more data or indirect blocks. Once a file’s inode has been
found, the number of disk I/Os required to read the file is
identical in Sprite LFS and Unix FFS.

Throughout this paper we use the Berkeley Unix fast
file system (Unix FFS) as an example of current file system
design and compare it to log-structured file systems. The
Unix FFS design is used because it is well documented in
the literature and used in several popular Unix operating
systems. The problems presented in this section are not
unique to Unix FFS and can be found in most other file systems.

3. Log-structured file systems
The fundamental idea of a log-structured file system
is to improve write performance by buffering a sequence of
file system changes in the file cache and then writing all the
changes to disk sequentially in a single disk write operation. The information written to disk in the write operation
includes file data blocks, attributes, index blocks,

In Unix FFS each inode is at a fixed location on disk;
given the identifying number for a file, a simple calculation



















































































































 Location  Section 
Data structure
Purpose


















































































































 



















































































































Inode
Locates blocks of file, holds protection bits, modify time, etc.
Log
3.1 

 


















































































































Inode map
Log
3.1 

 















































































































 Locates position of inode in log, holds time of last access plus version number.


Indirect block
 Locates blocks of large files.

Log

3.1 

 















































































































 Identifies contents of segment (file number and offset for each block).


Segment summary
Log
3.2 

 


















































































































Segment usage table  Counts live bytes still left in segments, stores last write time for data in segments.
Log
3.6 

 

















































































































Superblock

 















































































































 Holds static configuration information such as number of segments and segment size.
 Fixed  None 

















































































































Checkpoint region
 Locates blocks of inode map and segment usage table, identifies last checkpoint in log.  Fixed 
4.1 



















































































































Directory change log  Records directory operations to maintain consistency of reference counts in inodes.
Log
4.2 

Table 1 — Summary of the major data structures stored on disk by Sprite LFS.
For each data structure the table indicates the purpose served by the data structure in Sprite LFS. The table also indicates whether the data
structure is stored in the log or at a fixed position on disk and where in the paper the data structure is discussed in detail. Inodes, indirect
blocks, and superblocks are similar to the Unix FFS data structures with the same names. Note that Sprite LFS contains neither a bitmap
nor a free list.

July 24, 1991

-3-

yields the disk address of the file’s inode. In contrast,
Sprite LFS doesn’t place inodes at fixed positions; they are
written to the log. Sprite LFS uses a data structure called
an inode map to maintain the current location of each
inode. Given the identifying number for a file, the inode
map must be indexed to determine the disk address of the
inode. The inode map is divided into blocks that are written to the log; a fixed checkpoint region on each disk
identifies the locations of all the inode map blocks. Fortunately, inode maps are compact enough to keep the active
portions cached in main memory: inode map lookups
rarely require disk accesses.

traditional file systems. The second alternative is to copy
live data out of the log in order to leave large free extents
for writing. For this paper we will assume that the live data
is written back in a compacted form at the head of the log;
it could also be moved to another log-structured file system
to form a hierarchy of logs, or it could be moved to some
totally different file system or archive. The disadvantage of
copying is its cost, particularly for long-lived files; in the
simplest case where the log works circularly across the disk
and live data is copied back into the log, all of the longlived files will have to be copied in every pass of the log
across the disk.

Figure 1 shows the disk layouts that would occur in
Sprite LFS and Unix FFS after creating two new files in
different directories. Although the two layouts have the
same logical structure, the log-structured file system produces a much more compact arrangement. As a result, the
write performance of Sprite LFS is much better than Unix
FFS, while its read performance is just as good.

Sprite LFS uses a combination of threading and
copying. The disk is divided into large fixed-size extents
called segments. Any given segment is always written
sequentially from its beginning to its end, and all live data
must be copied out of a segment before the segment can be
rewritten. However, the log is threaded on a segment-bysegment basis; if the system can collect long-lived data
together into segments, those segments can be skipped over
so that the data doesn’t have to be copied repeatedly. The
segment size is chosen large enough that the transfer time
to read or write a whole segment is much greater than the
cost of a seek to the beginning of the segment. This allows
whole-segment operations to run at nearly the full
bandwidth of the disk, regardless of the order in which segments are accessed. Sprite LFS currently uses segment
sizes of either 512 kilobytes or one megabyte.

3.2. Free space management: segments
The most difficult design issue for log-structured file
systems is the management of free space. The goal is to
maintain large free extents for writing new data. Initially
all the free space is in a single extent on disk, but by the
time the log reaches the end of the disk the free space will
have been fragmented into many small extents corresponding to the files that were deleted or overwritten.
From this point on, the file system has two choices:
threading and copying. These are illustrated in Figure 2.
The first alternative is to leave the live data in place and
thread the log through the free extents. Unfortunately,
threading will cause the free space to become severely
fragmented, so that large contiguous writes won’t be possible and a log-structured file system will be no faster than

dir1

The process of copying live data out of a segment is
called segment cleaning. In Sprite LFS it is a simple
three-step process: read a number of segments into
memory, identify the live data, and write the live data back
to a smaller number of clean segments. After this

dir2

file1
Log

file1

3.3. Segment cleaning mechanism

Disk

Disk

Sprite LFS

file2
Block key:

file2

Inode

dir1

Directory

Data

dir2

Unix FFS

Inode map

Figure 1 — A comparison between Sprite LFS and Unix FFS.
This example shows the modified disk blocks written by Sprite LFS and Unix FFS when creating two single-block files named
dir1/file1 and dir2/file2. Each system must write new data blocks and inodes for file1 and file2, plus new data blocks
and inodes for the containing directories. Unix FFS requires ten non-sequential writes for the new information (the inodes for the new files
are each written twice to ease recovery from crashes), while Sprite LFS performs the operations in a single large write. The same number
of disk accesses will be required to read the files in the two systems. Sprite LFS also writes out new inode map blocks to record the new
inode locations.

July 24, 1991

-4-

operation is complete, the segments that were read are
marked as clean, and they can be used for new data or for
additional cleaning.

the segment; if the uid of a block does not match the uid
currently stored in the inode map when the segment is
cleaned, the block can be discarded immediately without
examining the file’s inode.

As part of segment cleaning it must be possible to
identify which blocks of each segment are live, so that they
can be written out again. It must also be possible to identify the file to which each block belongs and the position of
the block within the file; this information is needed in order
to update the file’s inode to point to the new location of the
block. Sprite LFS solves both of these problems by writing
a segment summary block as part of each segment. The
summary block identifies each piece of information that is
written in the segment; for example, for each file data block
the summary block contains the file number and block
number for the block. Segments can contain multiple segment summary blocks when more than one log write is
needed to fill the segment. (Partial-segment writes occur
when the number of dirty blocks buffered in the file cache
is insufficient to fill a segment.) Segment summary blocks
impose little overhead during writing, and they are useful
during crash recovery (see Section 4) as well as during
cleaning.

This approach to cleaning means that there is no
free-block list or bitmap in Sprite. In addition to saving
memory and disk space, the elimination of these data structures also simplifies crash recovery. If these data structures
existed, additional code would be needed to log changes to
the structures and restore consistency after crashes.

3.4. Segment cleaning policies
Given the basic mechanism described above, four
policy issues must be addressed:

Sprite LFS also uses the segment summary information to distinguish live blocks from those that have been
overwritten or deleted. Once a block’s identity is known,
its liveness can be determined by checking the file’s inode
or indirect block to see if the appropriate block pointer still
refers to this block. If it does, then the block is live; if it
doesn’t, then the block is dead. Sprite LFS optimizes this
check slightly by keeping a version number in the inode
map entry for each file; the version number is incremented
whenever the file is deleted or truncated to length zero.
The version number combined with the inode number form
an unique identifier (uid) for the contents of the file. The
segment summary block records this uid for each block in

Block Key:
Old data block

When should the segment cleaner execute? Some
possible choices are for it to run continuously in
background at low priority, or only at night, or only
when disk space is nearly exhausted.

(2)

How many segments should it clean at a time? Segment cleaning offers an opportunity to reorganize
data on disk; the more segments cleaned at once, the
more opportunities to rearrange.

(3)

Which segments should be cleaned? An obvious
choice is the ones that are most fragmented, but this
turns out not to be the best choice.

(4)

How should the live blocks be grouped when they
are written out? One possibility is to try to enhance
the locality of future reads, for example by grouping
files in the same directory together into a single output segment. Another possibility is to sort the blocks
by the time they were last modified and group blocks
of similar age together into new segments; we call
this approach age sort.

Copy and Compact

Threaded log
Old log end

(1)

New log end

Old log end

New log end

New data block
Previously deleted

Figure 2 — Possible free space management solutions for log-structured file systems.
In a log-structured file system, free space for the log can be generated either by copying the old blocks or by threading the log around the
old blocks. The left side of the figure shows the threaded log approach where the log skips over the active blocks and overwrites blocks of
files that have been deleted or overwritten. Pointers between the blocks of the log are maintained so that the log can be followed during
crash recovery. The right side of the figure shows the copying scheme where log space is generated by reading the section of disk after the
end of the log and rewriting the active blocks of that section along with the new data into the newly generated space.

July 24, 1991

-5-

In our work so far we have not methodically
addressed the first two of the above policies. Sprite LFS
starts cleaning segments when the number of clean segments drops below a threshold value (typically a few tens
of segments). It cleans a few tens of segments at a time
until the number of clean segments surpasses another threshold value (typically 50-100 clean segments). The overall
performance of Sprite LFS does not seem to be very sensitive to the exact choice of the threshold values. In contrast,
the third and fourth policy decisions are critically important: in our experience they are the primary factors that
determine the performance of a log-structured file system.
The remainder of Section 3 discusses our analysis of which
segments to clean and how to group the live data.

Figure 3 graphs the write cost as a function of u. For
reference, Unix FFS on small-file workloads utilizes at
most 5-10% of the disk bandwidth, for a write cost of
10-20 (see [11] and Figure 8 in Section 5.1 for specific
measurements). With logging, delayed writes, and disk
request sorting this can probably be improved to about 25%
of the bandwidth[12] or a write cost of 4. Figure 3 suggests
that the segments cleaned must have a utilization of less
than .8 in order for a log-structured file system to outperform the current Unix FFS; the utilization must be less than
.5 to outperform an improved Unix FFS.
It is important to note that the utilization discussed
above is not the overall fraction of the disk containing live
data; it is just the fraction of live blocks in segments that
are cleaned. Variations in file usage will cause some segments to be less utilized than others, and the cleaner can
choose the least utilized segments to clean; these will have
lower utilization than the overall average for the disk.

We use a term called write cost to compare cleaning
policies. The write cost is the average amount of time the
disk is busy per byte of new data written, including all the
cleaning overheads. The write cost is expressed as a multiple of the time that would be required if there were no
cleaning overhead and the data could be written at its full
bandwidth with no seek time or rotational latency. A write
cost of 1.0 is perfect: it would mean that new data could be
written at the full disk bandwidth and there is no cleaning
overhead. A write cost of 10 means that only one-tenth of
the disk’s maximum bandwidth is actually used for writing
new data; the rest of the disk time is spent in seeks, rotational latency, or cleaning.

Even so, the performance of a log-structured file system can be improved by reducing the overall utilization of
the disk space. With less of the disk in use the segments
that are cleaned will have fewer live blocks resulting in a
lower write cost. Log-structured file systems provide a
cost-performance tradeoff: if disk space is underutilized,
higher performance can be achieved but at a high cost per
usable byte; if disk capacity utilization is increased, storage
costs are reduced but so is performance. Such a tradeoff

For a log-structured file system with large segments,
seeks and rotational latency are negligible both for writing
and for cleaning, so the write cost is the total number of
bytes moved to and from the disk divided by the number of
those bytes that represent new data. This cost is determined by the utilization (the fraction of data still live) in
the segments that are cleaned. In the steady state, the
cleaner must generate one clean segment for every segment
of new data written. To do this, it reads N segments in
their entirety and writes out N*u segments of live data
(where u is the utilization of the segments and 0 ≤ u < 1).
This creates N*(1−u) segments of contiguous free space for
new data. Thus
write cost =

Write cost
14.0
10.0

read segs + write live + write new
new data written

=

N + N*u + N*(1−u)
2
=
N*(1−u)
1−u

6.0
4.0

FFS improved

2.0
0.0

0.0

0.2

0.4

0.6

0.8

1.0

Fraction alive in segment cleaned (u)
Figure 3 — Write cost as a function of u for small files.
In a log-structured file system, the write cost depends strongly on
the utilization of the segments that are cleaned. The more live
data in segments cleaned the more disk bandwidth that is needed
for cleaning and not available for writing new data. The figure
also shows two reference points: ‘‘FFS today’’, which represents
Unix FFS today, and ‘‘FFS improved’’, which is our estimate of
the best performance possible in an improved Unix FFS. Write
cost for Unix FFS is not sensitive to the amount of disk space in
use.

(1)

In the above formula we made the conservative assumption
that a segment must be read in its entirety to recover the
live blocks; in practice it may be faster to read just the live
blocks, particularly if the utilization is very low (we
haven’t tried this in Sprite LFS). If a segment to be cleaned
has no live blocks (u = 0) then it need not be read at all and
the write cost is 1.0.
July 24, 1991

FFS today

8.0

total bytes read and written
new data written

=

Log-structured

12.0

-6-

between performance and space utilization is not unique to
log-structured file systems. For example, Unix FFS only
allows 90% of the disk space to be occupied by files. The
remaining 10% is kept free to allow the space allocation
algorithm to operate efficiently.

Write cost
14.0
12.0
10.0
8.0
6.0
4.0
2.0
0.0

The key to achieving high performance at low cost in
a log-structured file system is to force the disk into a bimodal segment distribution where most of the segments are
nearly full, a few are empty or nearly empty, and the
cleaner can almost always work with the empty segments.
This allows a high overall disk capacity utilization yet provides a low write cost. The following section describes
how we achieve such a bimodal distribution in Sprite LFS.

3.5. Simulation results

Each file has equal likelihood of being
selected in each step.

Hot-and-cold

Files are divided into two groups. One
group contains 10% of the files; it is
called hot because its files are selected
90% of the time. The other group is
called cold; it contains 90% of the files
but they are selected only 10% of the
time. Within groups each file is equally
likely to be selected. This access pattern
models a simple form of locality.

FFS today
LFS uniform
FFS improved
0.0

0.2

0.4

0.6

0.8

1.0

Figure 4 — Initial simulation results.
The curves labeled ‘‘FFS today’’ and ‘‘FFS improved’’ are reproduced from Figure 3 for comparison. The curve labeled ‘‘No
variance’’ shows the write cost that would occur if all segments
always had exactly the same utilization. The ‘‘LFS uniform’’
curve represents a log-structured file system with uniform access
pattern and a greedy cleaning policy: the cleaner chooses the
least-utilized segments. The ‘‘LFS hot-and-cold’’ curve
represents a log-structured file system with locality of file access.
It uses a greedy cleaning policy and the cleaner also sorts the live
data by age before writing it out again. The x-axis is overall disk
capacity utilization, which is not necessarily the same as the utilization of the segments being cleaned.

Even with uniform random access patterns, the variance in segment utilization allows a substantially lower
write cost than would be predicted from the overall disk
capacity utilization and formula (1). For example, at 75%
overall disk capacity utilization, the segments cleaned have
an average utilization of only 55%. At overall disk capacity utilizations under 20% the write cost drops below 2.0;
this means that some of the cleaned segments have no live
blocks at all and hence don’t need to be read in.
The ‘‘LFS hot-and-cold’’ curve shows the write cost
when there is locality in the access patterns, as described
above. The cleaning policy for this curve was the same as
for ‘‘LFS uniform’’ except that the live blocks were sorted
by age before writing them out again. This means that
long-lived (cold) data tends to be segregated in different
segments from short-lived (hot) data; we thought that this
approach would lead to the desired bimodal distribution of
segment utilizations.

In this approach the overall disk capacity utilization is constant and no read traffic is modeled. The simulator runs
until all clean segments are exhausted, then simulates the
actions of a cleaner until a threshold number of clean segments is available again. In each run the simulator was
allowed to run until the write cost stabilized and all coldstart variance had been removed.
Figure 4 superimposes the results from two sets of
simulations onto the curves of Figure 3. In the ‘‘LFS uniform’’ simulations the uniform access pattern was used.
The cleaner used a simple greedy policy where it always
chose the least-utilized segments to clean. When writing
out live data the cleaner did not attempt to re-organize the
data: live blocks were written out in the same order that
they appeared in the segments being cleaned (for a uniform
access pattern there is no reason to expect any improvement from re-organization).
July 24, 1991

LFS hot-and-cold

Disk capacity utilization

We built a simple file system simulator so that we
could analyze different cleaning policies under controlled
conditions. The simulator’s model does not reflect actual
file system usage patterns (its model is much harsher than
reality), but it helped us to understand the effects of random access patterns and locality, both of which can be
exploited to reduce the cost of cleaning. The simulator
models a file system as a fixed number of 4-kbyte files,
with the number chosen to produce a particular overall disk
capacity utilization. At each step, the simulator overwrites
one of the files with new data, using one of two pseudorandom access patterns:
Uniform

No variance

Figure 4 shows the surprising result that locality and
‘‘better’’ grouping result in worse performance than a system with no locality! We tried varying the degree of locality (e.g. 95% of accesses to 5% of data) and found that performance got worse and worse as the locality increased.
Figure 5 shows the reason for this non-intuitive result.
Under the greedy policy, a segment doesn’t get cleaned
until it becomes the least utilized of all segments. Thus
every segment’s utilization eventually drops to the cleaning
threshold, including the cold segments. Unfortunately, the
-7-

remain unchanged, the stability can be estimated by the age
of data.

Fraction of segments
0.008
0.007
0.006
0.005
0.004
0.003
0.002
0.001
0.000

To test this theory we simulated a new policy for
selecting segments to clean. The policy rates each segment
according to the benefit of cleaning the segment and the
cost of cleaning the segment and chooses the segments with
the highest ratio of benefit to cost. The benefit has two
components: the amount of free space that will be
reclaimed and the amount of time the space is likely to stay
free. The amount of free space is just 1−u, where u is the
utilization of the segment. We used the most recent
modified time of any block in the segment (ie. the age of
the youngest block) as an estimate of how long the space is
likely to stay free. The benefit of cleaning is the space-time
product formed by multiplying these two components. The
cost of cleaning the segment is 1+u (one unit of cost to read
the segment, u to write back the live data). Combining all
these factors, we get

Hot-and-cold
Uniform

0.0

0.2

0.4

0.6

0.8

1.0

Segment utilization
Figure 5 — Segment utilization distributions with greedy
cleaner.
These figures show distributions of segment utilizations of the
disk during the simulation. The distribution is computed by
measuring the utilizations of all segments on the disk at the points
during the simulation when segment cleaning was initiated. The
distribution shows the utilizations of the segments available to the
cleaning algorithm. Each of the distributions corresponds to an
overall disk capacity utilization of 75%. The ‘‘Uniform’’ curve
corresponds to ‘‘LFS uniform’’ in Figure 4 and ‘‘Hot-and-cold’’
corresponds to ‘‘LFS hot-and-cold’’ in Figure 4. Locality causes
the distribution to be more skewed towards the utilization at
which cleaning occurs; as a result, segments are cleaned at a
higher average utilization.

benefit
free space generated * age of data
(1−u)*age
=
=
cost
cost
1+u
We call this policy the cost-benefit policy; it allows cold
segments to be cleaned at a much higher utilization than
hot segments.
We re-ran the simulations under the hot-and-cold
access pattern with the cost-benefit policy and age-sorting

Fraction of segments
0.008
0.007
0.006
0.005
0.004
0.003
0.002
0.001
0.000

utilization drops very slowly in cold segments, so these
segments tend to linger just above the cleaning point for a
very long time. Figure 5 shows that many more segments
are clustered around the cleaning point in the simulations
with locality than in the simulations without locality. The
overall result is that cold segments tend to tie up large
numbers of free blocks for long periods of time.
After studying these figures we realized that hot and
cold segments must be treated differently by the cleaner.
Free space in a cold segment is more valuable than free
space in a hot segment because once a cold segment has
been cleaned it will take a long time before it reaccumulates the unusable free space. Said another way,
once the system reclaims the free blocks from a segment
with cold data it will get to ‘‘keep’’ them a long time
before the cold data becomes fragmented and ‘‘takes them
back again.’’ In contrast, it is less beneficial to clean a hot
segment because the data will likely die quickly and the
free space will rapidly re-accumulate; the system might as
well delay the cleaning a while and let more of the blocks
die in the current segment. The value of a segment’s free
space is based on the stability of the data in the segment.
Unfortunately, the stability cannot be predicted without
knowing future access patterns. Using an assumption that
the older the data in a segment the longer it is likely to
July 24, 1991

LFS Cost-Benefit
LFS Greedy

0.0

0.2

0.4

0.6

0.8

1.0

Segment utilization
Figure 6 — Segment utilization distribution with cost-benefit
policy.
This figure shows the distribution of segment utilizations from the
simulation of a hot-and-cold access pattern with 75% overall disk
capacity utilization. The ‘‘LFS Cost-Benefit’’ curve shows the
segment distribution occurring when the cost-benefit policy is
used to select segments to clean and live blocks grouped by age
before being re-written. Because of this bimodal segment distribution, most of the segments cleaned had utilizations around 15%.
For comparison, the distribution produced by the greedy method
selection policy is shown by the ‘‘LFS Greedy’’ curve reproduced
from Figure 5.

-8-

on the live data. As can be seen from Figure 6, the costbenefit policy produced the bimodal distribution of segments that we had hoped for. The cleaning policy cleans
cold segments at about 75% utilization but waits until hot
segments reach a utilization of about 15% before cleaning
them. Since 90% of the writes are to hot files, most of the
segments cleaned are hot. Figure 7 shows that the costbenefit policy reduces the write cost by as much as 50%
over the greedy policy, and a log-structured file system
out-performs the best possible Unix FFS even at relatively
high disk capacity utilizations. We simulated a number of
other degrees and kinds of locality and found that the costbenefit policy gets even better as locality increases.

the checkpoint regions (see Section 4 for details).
In order to sort live blocks by age, the segment summary information records the age of the youngest block
written to the segment. At present Sprite LFS does not
keep modified times for each block in a file; it keeps a single modified time for the entire file. This estimate will be
incorrect for files that are not modified in their entirety.
We plan to modify the segment summary information to
include modified times for each block.

4. Crash recovery
When a system crash occurs, the last few operations
performed on the disk may have left it in an inconsistent
state (for example, a new file may have been written
without writing the directory containing the file); during
reboot the operating system must review these operations
in order to correct any inconsistencies. In traditional Unix
file systems without logs, the system cannot determine
where the last changes were made, so it must scan all of the
metadata structures on disk to restore consistency. The
cost of these scans is already high (tens of minutes in typical configurations), and it is getting higher as storage systems expand.

The simulation experiments convinced us to implement the cost-benefit approach in Sprite LFS. As will be
seen in Section 5.2, the behavior of actual file systems in
Sprite LFS is even better than predicted in Figure 7.

3.6. Segment usage table
In order to support the cost-benefit cleaning policy,
Sprite LFS maintains a data structure called the segment
usage table. For each segment, the table records the
number of live bytes in the segment and the most recent
modified time of any block in the segment. These two
values are used by the segment cleaner when choosing segments to clean. The values are initially set when the segment is written, and the count of live bytes is decremented
when files are deleted or blocks are overwritten. If the
count falls to zero then the segment can be reused without
cleaning. The blocks of the segment usage table are written to the log, and the addresses of the blocks are stored in

In a log-structured file system the locations of the last
disk operations are easy to determine: they are at the end
of the log. Thus it should be possible to recover very
quickly after crashes. This benefit of logs is well known
and has been used to advantage both in database systems[13] and in other file systems[2, 3, 14]. Like many
other logging systems, Sprite LFS uses a two-pronged
approach to recovery: checkpoints, which define consistent
states of the file system, and roll-forward, which is used to
recover information written since the last checkpoint.

Write cost
14.0

4.1. Checkpoints

No variance

12.0

A checkpoint is a position in the log at which all of
the file system structures are consistent and complete.
Sprite LFS uses a two-phase process to create a checkpoint.
First, it writes out all modified information to the log,
including file data blocks, indirect blocks, inodes, and
blocks of the inode map and segment usage table. Second,
it writes a checkpoint region to a special fixed position on
disk. The checkpoint region contains the addresses of all
the blocks in the inode map and segment usage table, plus
the current time and a pointer to the last segment written.

LFS Greedy

10.0

FFS today

8.0
6.0

LFS Cost-Benefit

4.0

FFS improved

2.0
0.0

0.0

0.2

0.4

0.6

0.8

1.0

During reboot, Sprite LFS reads the checkpoint
region and uses that information to initialize its mainmemory data structures. In order to handle a crash during a
checkpoint operation there are actually two checkpoint
regions, and checkpoint operations alternate between them.
The checkpoint time is in the last block of the checkpoint
region, so if the checkpoint fails the time will not be
updated. During reboot, the system reads both checkpoint
regions and uses the one with the most recent time.

Disk capacity utilization
Figure 7 — Write cost, including cost-benefit policy.
This graph compares the write cost of the greedy policy with that
of the cost-benefit policy for the hot-and-cold access pattern. The
cost-benefit policy is substantially better than the greedy policy,
particularly for disk capacity utilizations above 60%.

Sprite LFS performs checkpoints at periodic intervals
as well as when the file system is unmounted or the system
July 24, 1991

-9-

is shut down. A long interval between checkpoints reduces
the overhead of writing the checkpoints but increases the
time needed to roll forward during recovery; a short
checkpoint interval improves recovery time but increases
the cost of normal operation. Sprite LFS currently uses a
checkpoint interval of thirty seconds, which is probably
much too short. An alternative to periodic checkpointing is
to perform checkpoints after a given amount of new data
has been written to the log; this would set a limit on
recovery time while reducing the checkpoint overhead
when the file system is not operating at maximum
throughput.

4.2. Roll-forward
In principle it would be safe to restart after crashes
by simply reading the latest checkpoint region and discarding any data in the log after that checkpoint. This would
result in instantaneous recovery but any data written since
the last checkpoint would be lost. In order to recover as
much information as possible, Sprite LFS scans through the
log segments that were written after the last checkpoint.
This operation is called roll-forward.
During roll-forward Sprite LFS uses the information
in segment summary blocks to recover recently-written file
data. When a summary block indicates the presence of a
new inode, Sprite LFS updates the inode map it read from
the checkpoint, so that the inode map refers to the new
copy of the inode. This automatically incorporates the
file’s new data blocks into the recovered file system. If
data blocks are discovered for a file without a new copy of
the file’s inode, then the roll-forward code assumes that the
new version of the file on disk is incomplete and it ignores
the new data blocks.
The roll-forward code also adjusts the utilizations in
the segment usage table read from the checkpoint. The
utilizations of the segments written since the checkpoint
will be zero; they must be adjusted to reflect the live data
left after roll-forward. The utilizations of older segments
will also have to be adjusted to reflect file deletions and
overwrites (both of these can be identified by the presence
of new inodes in the log).
The final issue in roll-forward is how to restore consistency between directory entries and inodes. Each inode
contains a count of the number of directory entries referring to that inode; when the count drops to zero the file is
deleted. Unfortunately, it is possible for a crash to occur
when an inode has been written to the log with a new reference count while the block containing the corresponding
directory entry has not yet been written, or vice versa.

called the directory operation log; Sprite LFS guarantees
that each directory operation log entry appears in the log
before the corresponding directory block or inode.
During roll-forward, the directory operation log is
used to ensure consistency between directory entries and
inodes: if a log entry appears but the inode and directory
block were not both written, roll-forward updates the directory and/or inode to complete the operation. Roll-forward
operations can cause entries to be added to or removed
from directories and reference counts on inodes to be
updated. The recovery program appends the changed directories, inodes, inode map, and segment usage table blocks
to the log and writes a new checkpoint region to include
them. The only operation that can’t be completed is the
creation of a new file for which the inode is never written;
in this case the directory entry will be removed. In addition
to its other functions, the directory log made it easy to provide an atomic rename operation.
The interaction between the directory operation log
and checkpoints introduced additional synchronization
issues into Sprite LFS. In particular, each checkpoint must
represent a state where the directory operation log is consistent with the inode and directory blocks in the log. This
required additional synchronization to prevent directory
modifications while checkpoints are being written.

5. Experience with the Sprite LFS
We began the implementation of Sprite LFS in late
1989 and by mid-1990 it was operational as part of the
Sprite network operating system. Since the fall of 1990 it
has been used to manage five different disk partitions,
which are used by about thirty users for day-to-day computing. All of the features described in this paper have
been implemented in Sprite LFS, but roll-forward has not
yet been installed in the production system. The production disks use a short checkpoint interval (30 seconds) and
discard all the information after the last checkpoint when
they reboot.
When we began the project we were concerned that a
log-structured file system might be substantially more complicated to implement than a traditional file system. In reality, however, Sprite LFS turns out to be no more complicated than Unix FFS[9]: Sprite LFS has additional complexity for the segment cleaner, but this is compensated by
the elimination of the bitmap and layout policies required
by Unix FFS; in addition, the checkpointing and rollforward code in Sprite LFS is no more complicated than
the fsck code[15] that scans Unix FFS disks to restore consistency. Logging file systems like Episode[2] or Cedar[3]
are likely to be somewhat more complicated than either
Unix FFS or Sprite LFS, since they include both logging
and layout code.

To restore consistency between directories and
inodes, Sprite LFS outputs a special record in the log for
each directory change. The record includes an operation
code (create, link, rename, or unlink), the location of the
In everyday use Sprite LFS does not feel much difdirectory entry (i-number for the directory and the position
ferent to the users than the Unix FFS-like file system in
within the directory), the contents of the directory entry
Sprite. The reason is that the machines being used are not
(name and i-number), and the new reference count for the
fast enough to be disk-bound with the current workloads.
inode named in the entry. These records are collectively
For example on the modified Andrew benchmark[11],
July 24, 1991
- 10 -

running multiuser but was otherwise quiescent during the
test. For Sprite LFS no cleaning occurred during the
benchmark runs so the measurements represent best-case
performance; see Section 5.2 below for measurements of
cleaning overhead.

Key:
Sprite LFS
SunOS
Files/sec (measured)
Files/sec (predicted)

180
160
140
120
100
80
60
40
20
0

Create Read Delete
10000 1K file access
(a)

675
600
525
450
375
300
225
150
75
0

Figure 8 shows the results of a benchmark that
creates, reads, and deletes a large number of small files.
Sprite LFS is almost ten times as fast as SunOS for the
create and delete phases of the benchmark. Sprite LFS is
also faster for reading the files back; this is because the
files are read in the same order created and the logstructured file system packs the files densely in the log.
Furthermore, Sprite LFS only kept the disk 17% busy during the create phase while saturating the CPU. In contrast,
SunOS kept the disk busy 85% of the time during the create
phase, even though only about 1.2% of the disk’s potential
bandwidth was used for new data. This means that the performance of Sprite LFS will improve by another factor of
4-6 as CPUs get faster (see Figure 8(b)). Almost no
improvement can be expected in SunOS.

Sun4 2*Sun4 4*Sun4
10000 1K file create
(b)

Figure 8 — Small-file performance under Sprite LFS and
SunOS.
Figure (a) measures a benchmark that created 10000 one-kilobyte
files, then read them back in the same order as created, then deleted them. Speed is measured by the number of files per second for
each operation on the two file systems. The logging approach in
Sprite LFS provides an order-of-magnitude speedup for creation
and deletion. Figure (b) estimates the performance of each system
for creating files on faster computers with the same disk. In
SunOS the disk was 85% saturated in (a), so faster processors will
not improve performance much. In Sprite LFS the disk was only
17% saturated in (a) while the CPU was 100% utilized; as a
consequence I/O performance will scale with CPU speed.

Although Sprite was designed for efficiency on
workloads with many small file accesses, Figure 9 shows
that it also provides competitive performance for large
files. Sprite LFS has a higher write bandwidth than SunOS
in all cases. It is substantially faster for random writes
because it turns them into sequential writes to the log; it is
also faster for sequential writes because it groups many
blocks into a single large I/O, whereas SunOS performs

kilobytes/sec
900
800
700
600
500
400
300
200
100
0

Sprite LFS is only 20% faster that SunOS using the
configuration presented in Section 5.1. Most of the
speedup is attributable to the removal of the synchronous
writes in Sprite LFS. Even with the synchronous writes of
Unix FFS, the benchmark has a CPU utilization of over
80%, limiting the speedup possible from changes in the
disk storage management.

5.1. Micro-benchmarks
We used a collection of small benchmark programs
to measure the best-case performance of Sprite LFS and
compare it to SunOS 4.0.3, whose file system is based on
Unix FFS. The benchmarks are synthetic so they do not
represent realistic workloads, but they illustrate the
strengths and weaknesses of the two file systems. The
machine used for both systems was a Sun-4/260 (8.7
integer SPECmarks) with 32 megabytes of memory, a Sun
SCSI3 HBA, and a Wren IV disk (1.3 MBytes/sec maximum transfer bandwidth, 17.5 milliseconds average seek
time). For both LFS and SunOS, the disk was formatted
with a file system having around 300 megabytes of usable
storage. An eight-kilobyte block size was used by SunOS
while Sprite LFS used a four-kilobyte block size and a
one-megabyte segment size. In each case the system was
July 24, 1991

Sprite LFS

Write
Read
Sequential

Write
Read
Random

SunOS

Reread
Sequential

Figure 9 — Large-file performance under Sprite LFS and
SunOS.
The figure shows the speed of a benchmark that creates a 100Mbyte file with sequential writes, then reads the file back sequentially, then writes 100 Mbytes randomly to the existing file, then
reads 100 Mbytes randomly from the file, and finally reads the file
sequentially again. The bandwidth of each of the five phases is
shown separately. Sprite LFS has a higher write bandwidth and
the same read bandwidth as SunOS with the exception of sequential reading of a file that was written randomly.

- 11 -

individual disk operations for each block (a newer version
of SunOS groups writes [16] and should therefore have
performance equivalent to Sprite LFS). The read performance is similar in the two systems except for the case of
reading a file sequentially after it has been written randomly; in this case the reads require seeks in Sprite LFS, so
its performance is substantially lower than SunOS.

write cost during the benchmark runs was 1.0). In order to
assess the cost of cleaning and the effectiveness of the
cost-benefit cleaning policy, we recorded statistics about
our production log-structured file systems over a period of
several months. Five systems were measured:
/user6 Home directories for Sprite developers. Workload consists of program development, text processing, electronic communication, and simulations.

Figure 9 illustrates the fact that a log-structured file
system produces a different form of locality on disk than
traditional file systems. A traditional file system achieves
logical locality by assuming certain access patterns
(sequential reading of files, a tendency to use multiple files
within a directory, etc.); it then pays extra on writes, if
necessary, to organize information optimally on disk for the
assumed read patterns. In contrast, a log-structured file
system achieves temporal locality: information that is
created or modified at the same time will be grouped
closely on disk. If temporal locality matches logical locality, as it does for a file that is written sequentially and then
read sequentially, then a log-structured file system should
have about the same performance on large files as a traditional file system. If temporal locality differs from logical
locality then the systems will perform differently. Sprite
LFS handles random writes more efficiently because it
writes them sequentially on disk. SunOS pays more for the
random writes in order to achieve logical locality, but then
it handles sequential re-reads more efficiently. Random
reads have about the same performance in the two systems,
even though the blocks are laid out very differently. However, if the nonsequential reads occurred in the same order
as the nonsequential writes then Sprite would have been
much faster.

/pcs

Home directories and project area for research on
parallel processing and VLSI circuit design.

/src/kernel
Sources and binaries for the Sprite kernel.
/swap2 Sprite client workstation swap files. Workload
consists of virtual memory backing store for 40
diskless Sprite workstations. Files tend to be
large, sparse, and accessed nonsequentially.
/tmp

Temporary file storage area for 40 Sprite workstations.

Table 2 shows statistics gathered during cleaning
over a four-month period. In order to eliminate start-up
effects we waited several months after putting the file systems into use before beginning the measurements. The
behavior of the production file systems has been substantially better than predicted by the simulations in Section 3.
Even though the overall disk capacity utilizations ranged
from 11-75%, more than half of the segments cleaned were
totally empty. Even the non-empty segments have utilizations far less than the average disk utilizations. The overall
write costs ranged from 1.2 to 1.6, in comparison to write
costs of 2.5-3 in the corresponding simulations. Figure 10
shows the distribution of segment utilizations, gathered in a
recent snapshot of the /user6 disk.

5.2. Cleaning overheads
The micro-benchmark results of the previous section
give an optimistic view of the performance of Sprite LFS
because they do not include any cleaning overheads (the

We believe that there are two reasons why cleaning
costs are lower in Sprite LFS than in the simulations. First,
































































































































































































Write cost in Sprite LFS file systems


































































































 Avg File 


 u  Write
Disk
Avg Write
Segments
File system
In Use 








Size
Size
Traffic
Cleaned  Empty
Avg
Cost

































































































 1280 MB  23.5 KB  3.2 MB/hour  75%  10732  69%  .133  1.4
 /user6








 52%
990
MB
10.5
KB
2.1
MB/hour
63%
22689
.137
1.6
 /pcs









 /src/kernel  1280 MB  37.5 KB  4.2 MB/hour  72%  16975
83%
.122
1.2




 /tmp

264
MB
28.9
KB
1.7
MB/hour
11%
2871
78%
.130
1.3








































































































4701  66%  .535  1.6
 /swap2
 309 MB  68.1 KB  13.3 MB/hour  65% 
Table 2 - Segment cleaning statistics and write costs for production file systems.
For each Sprite LFS file system the table lists the disk size, the average file size, the average daily write traffic rate, the average disk capacity utilization, the total number of segments cleaned over a four-month period, the fraction of the segments that were empty when cleaned,
the average utilization of the non-empty segments that were cleaned, and the overall write cost for the period of the measurements. These
write cost figures imply that the cleaning overhead limits the long-term write performance to about 70% of the maximum sequential write
bandwidth.

July 24, 1991

- 12 -

all the files in the simulations were just a single block long.
In practice, there are a substantial number of longer files,
and they tend to be written and deleted as a whole. This
results in greater locality within individual segments. In
the best case where a file is much longer than a segment,
deleting the file will produce one or more totally empty
segments. The second difference between simulation and
reality is that the simulated reference patterns were evenly
distributed within the hot and cold file groups. In practice
there are large numbers of files that are almost never written (cold segments in reality are much colder than the cold
segments in the simulations). A log-structured file system
will isolate the very cold files in segments and never clean
them. In the simulations, every segment eventually
received modifications and thus had to be cleaned.

5.3. Crash recovery

If the measurements of Sprite LFS in Section 5.1
were a bit over-optimistic, the measurements in this section
are, if anything, over-pessimistic. In practice it may be
possible to perform much of the cleaning at night or during
other idle periods, so that clean segments are available during bursts of activity. We do not yet have enough experience with Sprite LFS to know if this can be done. In addition, we expect the performance of Sprite LFS to improve
as we gain experience and tune the algorithms. For example, we have not yet carefully analyzed the policy issue of
how many segments to clean at a time, but we think it may
impact the system’s ability to segregate hot data from cold
data.

Table 3 shows that recovery time varies with the
number and size of files written between the last checkpoint and the crash. Recovery times can be bounded by
limiting the amount of data written between checkpoints.
From the average file sizes and daily write traffic in Table
2, a checkpoint interval as large as an hour would result in
average recovery times of around one second. Using the
maximum observered write rate of 150 megabytes/hour,
maximum recovery time would grow by one second for
every 70 seconds of checkpoint interval length.

Although the crash recovery code has not been
installed on the production system, the code works well
enough to time recovery of various crash scenarios. The
time to recover depends on the checkpoint interval and the
rate and type of operations being performed. Table 3
shows the recovery time for different file sizes and amounts
of file data recovered. The different crash configurations
were generated by running a program that created one, ten,
or fifty megabytes of fixed-size files before the system was
crashed. A special version of Sprite LFS was used that had
an infinite checkpoint interval and never wrote directory
changes to disk. During the recovery roll-forward, the
created files had to be added to the inode map, the directory
entries created, and the segment usage table updated.

5.4. Other overheads in Sprite LFS
Table 4 shows the relative importance of the various
kinds of data written to disk, both in terms of how much of
the live blocks they occupy on disk and in terms of how
much of the data written to the log they represent. More
than 99% of the live data on disk consists of file data
blocks and indirect blocks. However, about 13% of the
information written to the log consists of inodes, inode map
blocks, and segment map blocks, all of which tend to be
overwritten quickly. The inode map alone accounts for
more than 7% of all the data written to the log. We suspect
that this is because of the short checkpoint interval
currently used in Sprite LFS, which forces metadata to disk

Fraction of segments
0.180
0.160
0.140
0.120
0.100
0.080


 

































 
































Sprite LFS recovery time in seconds 
 File



























File Data Recovered





Size
1 MB 10 MB 50 MB 


 

































 132
1 KB 
1 
21




 10 KB  < 1
3
17




100
KB  < 1 
1
8 



































0.060
0.040
0.020
0.000

0.0

0.2

0.4

0.6

0.8

1.0

Segment utilization

Table 3 — Recovery time for various crash configurations
The table shows the speed of recovery of one, ten, and fifty megabytes of fixed-size files. The system measured was the same one
used in Section 5.1. Recovery time is dominated by the number of
files to be recovered.

Figure 10 — Segment utilization in the /user6 file system
This figure shows the distribution of segment utilizations in a recent snapshot of the /user6 disk. The distribution shows large
numbers of fully utilized segments and totally empty segments.

July 24, 1991

- 13 -

more often than necessary. We expect the log bandwidth
overhead for metadata to drop substantially when we install
roll-forward recovery and increase the checkpoint interval.

systems view the log as the most up to date ‘‘truth’’ about
the state of the data on disk. The main difference is that
database systems do not use the log as the final repository
for data: a separate data area is reserved for this purpose.
The separate data area of these database systems means
that they do not need the segment cleaning mechanisms of
the Sprite LFS to reclaim log space. The space occupied
by the log in a database system can be reclaimed when the
logged changes have been written to their final locations.
Since all read requests are processed from the data area, the
log can be greatly compacted without hurting read performance. Typically only the changed bytes are written to
database logs rather than entire blocks as in Sprite LFS.

6. Related work
The log-structured file system concept and the Sprite
LFS design borrow ideas from many different storage
management systems. File systems with log-like structures
have appeared in several proposals for building file systems
on write-once media[17, 18]. Besides writing all changes
in an append-only fashion, these systems maintain indexing
information much like the Sprite LFS inode map and
inodes for quickly locating and reading files. They differ
from Sprite LFS in that the write-once nature of the media
made it unnecessary for the file systems to reclaim log
space.

The Sprite LFS crash recovery mechanism of checkpoints and roll forward using a ‘‘redo log’’ is similar to
techniques used in database systems and object repositories[21]. The implementation in Sprite LFS is simplified
because the log is the final home of the data. Rather than
redoing the operation to the separate data copy, Sprite LFS
recovery insures that the indexes point at the newest copy
of the data in the log.

The segment cleaning approach used in Sprite LFS
acts much like scavenging garbage collectors developed for
programming languages[19]. The cost-benefit segment
selection and the age sorting of blocks during segment
cleaned in Sprite LFS separates files into generations much
like generational garbage collection schemes[20]. A
significant difference between these garbage collection
schemes and Sprite LFS is that efficient random access is
possible in the generational garbage collectors, whereas
sequential accesses are necessary to achieve high performance in a file system. Also, Sprite LFS can exploit the
fact that blocks can belong to at most one file at a time to
use much simpler algorithms for identifying garbage than
used in the systems for programming languages.

Collecting data in the file cache and writing it to disk
in large writes is similar to the concept of group commit in
database systems[22] and to techniques used in mainmemory database systems[23, 24].

7. Conclusion
The basic principle behind a log-structured file system is a simple one: collect large amounts of new data in a
file cache in main memory, then write the data to disk in a
single large I/O that can use all of the disk’s bandwidth.
Implementing this idea is complicated by the need to maintain large free areas on disk, but both our simulation
analysis and our experience with Sprite LFS suggest that
low cleaning overheads can be achieved with a simple policy based on cost and benefit. Although we developed a
log-structured file system to support workloads with many
small files, the approach also works very well for large-file
accesses. In particular, there is essentially no cleaning
overhead at all for very large files that are created and
deleted in their entirety.

The logging scheme used in Sprite LFS is similar to
schemes pioneered in database systems. Almost all database systems use write-ahead logging for crash recovery
and high performance[13], but differ from Sprite LFS in
how they use the log. Both Sprite LFS and the database























































































Sprite LFS /user6 file system contents















































Block
type
Live
data
Log
bandwidth













































 98.0% 
Data blocks*
85.2%














































Indirect blocks* 
1.0% 
1.6%
























































































Inode blocks*
0.2% 
2.7%

 Inode map


0.2% 
7.8%















































0.0%
2.1%











































 Seg Usage map* 


Summary blocks 
0.6% 
0.5%
























































































Dir Op Log
0.0%
0.1%



The bottom line is that a log-structured file system
can use disks an order of magnitude more efficiently than
existing file systems. This should make it possible to take
advantage of several more generations of faster processors
before I/O limitations once again threaten the scalability of
computer systems.

8. Acknowledgments

Table 4 — Disk space and log bandwidth usage of /user6
For each block type, the table lists the percentage of the disk
space in use on disk (Live data) and the percentage of the log
bandwidth consumed writing this block type (Log bandwidth).
The block types marked with ’*’ have equivalent data structures
in Unix FFS.

Diane Greene, Mary Baker, John Hartman, Mike
Kupfer, Ken Shirriff and Jim Mott-Smith provided helpful
comments on drafts of this paper.

References
1.

July 24, 1991

- 14 -

John K. Ousterhout, Herve Da Costa, David
Harrison, John A. Kunze, Mike Kupfer, and James

G. Thompson, ‘‘A Trace-Driven Analysis of the
Unix 4.2 BSD File System,’’ Proceedings of the 10th
Symposium on Operating Systems Principles, pp.
15-24 ACM, (1985).
2.

Michael L. Kazar, Bruce W. Leverett, Owen T.
Anderson, Vasilis Apostolides, Beth A. Bottos,
Sailesh Chutani, Craig F. Everhart, W. Anthony
Mason, Shu-Tsui Tu, and Edward R. Zayas,
‘‘DEcorum File System Architectural Overview,’’
Proceedings of the USENIX 1990 Summer Conference, pp. 151-164 (Jun 1990).

3.

Robert B. Hagmann, ‘‘Reimplementing the Cedar
File System Using Logging and Group Commit,’’
Proceedings of the 11th Symposium on Operating
Systems Principles, pp. 155-162 (Nov 1987).

4.

John K. Ousterhout, Andrew R. Cherenson, Frederick Douglis, Michael N. Nelson, and Brent B. Welch,
‘‘The Sprite Network Operating System,’’ IEEE
Computer 21(2) pp. 23-36 (1988).

5.

David A. Patterson, Garth Gibson, and Randy H.
Katz, ‘‘A Case for Redundant Arrays of Inexpensive
Disks (RAID),’’ ACM SIGMOD 88, pp. 109-116
(Jun 1988).

6.

Mary G. Baker, John H. Hartman, Michael D.
Kupfer, Ken W. Shirriff, and John K. Ousterhout,
‘‘Measurements of a Distributed File System,’’
Proceedings of the 13th Symposium on Operating
Systems Principles, ACM, (Oct 1991).

7.

8.

M. Satyanarayanan, ‘‘A Study of File Sizes and
Functional Lifetimes,’’ Proceedings of the 8th Symposium on Operating Systems Principles, pp. 96-108
ACM, (1981).
Edward D. Lazowska, John Zahorjan, David R Cheriton, and Willy Zwaenepoel, ‘‘File Access Performance of Diskless Workstations,’’ Transactions on
Computer Systems 4(3) pp. 238-268 (Aug 1986).

14.

A. Chang, M. F. Mergen, R. K. Rader, J. A. Roberts,
and S. L. Porter, ‘‘Evolution of storage facilities in
AIX Version 3 for RISC System/6000 processors,’’
IBM Journal of Research and Development 34(1) pp.
105-109 (Jan 1990).

15.

Marshall Kirk McKusick, Willian N. Joy, Samuel J.
Leffler, and Robert S. Fabry, ‘‘Fsck - The UNIX File
System Check Program,’’ Unix System Manager’s
Manual - 4.3 BSD Virtual VAX-11 Version,
USENIX, (Apr 1986).

16.

Larry McVoy and Steve Kleiman, ‘‘Extent-like Performance from a UNIX File System,’’ Proceedings
of the USENIX 1991 Winter Conference, (Jan 1991).

17.

D. Reed and Liba Svobodova, ‘‘SWALLOW: A Distributed Data Storage System for a Local Network,’’
Local Networks for Computer Communications, pp.
355-373 North-Holland, (1981).

18.

Ross S. Finlayson and David R. Cheriton, ‘‘Log
Files: An Extended File Service Exploiting WriteOnce Storage,’’ Proceedings of the 11th Symposium
on Operating Systems Principles, pp. 129-148 ACM,
(Nov 1987).

19.

H. G. Baker, ‘‘List Processing in Real Time on a
Serial Computer,’’ A.I. Working Paper 139, MIT-AI
Lab, Boston, MA (April 1977).

20.

Henry Lieberman and Carl Hewitt, ‘‘A Real-Time
Garbage Collector Based on the Lifetimes of
Objects,’’ Communications of the ACM 26(6) pp.
419-429 (1983).

21.

Brian M. Oki, Barbara H. Liskov, and Robert W.
Scheifler, ‘‘Reliable Object Storage to Support
Atomic Actions,’’ Proceedings of the 10th Symposium on Operating Systems Principles, pp. 147-159
ACM, (1985).

22.

David J. DeWitt, Randy H. Katz, Frank Olken, L. D.
Shapiro, Mike R. Stonebraker, and David Wood,
‘‘Implementation Techniques for Main Memory
Database Systems,’’ Proceedings of SIGMOD 1984,
pp. 1-8 (Jun 1984).

9.

Marshall K. McKusick, ‘‘A Fast File System for
Unix,’’ Transactions on Computer Systems 2(3) pp.
181-197 ACM, (1984).

10.

R. Sandberg, ‘‘Design and Implementation of the
Sun Network Filesystem,’’ Proceedings of the
USENIX 1985 Summer Conference, pp. 119-130
(Jun 1985).

23.

Kenneth Salem and Hector Garcia-Molina, ‘‘Crash
Recovery Mechanisms for Main Storage Database
Systems,’’ CS-TR-034-86, Princeton University,
Princeton, NJ (1986).

11.

John K. Ousterhout, ‘‘Why Aren’t Operating Systems Getting Faster As Fast as Hardware?,’’
Proceedings of the USENIX 1990 Summer Conference, pp. 247-256 (Jun 1990).

24.

Robert B. Hagmann, ‘‘A Crash Recovery Scheme for
a Memory-Resident Database System,’’ IEEE Transactions on Computers C-35(9)(Sep 1986).

12.

Margo I. Seltzer, Peter M. Chen, and John K.
Ousterhout, ‘‘Disk Scheduling Revisited,’’ Proceedings of the Winter 1990 USENIX Technical Conference, (January 1990).

13.

Jim Gray, ‘‘Notes on Data Base Operating Systems,’’ in Operating Systems, An Advanced Course,
Springer-Verlag (1979).

July 24, 1991

- 15 -