USENIX Technical Program - Paper - Proceedings of the USENIX Annual Technical Conference, January 6-10, 1997, Anaheim, California, USA
|Pp. 7790 of the Proceedings|
Extending the Operating System at the User Level:
the Ufo Global File System
Albert D. Alexandrov, Maximilian Ibel,
Klaus E. Schauser, and Chris J. Scheiman
Department of Computer Science
University of California, Santa Barbara
Santa Barbara, CA 93106
In this paper we show how to extend the functionality of
standard operating systems completely at the user level. Our approach works by
intercepting selected system calls at the user level, using
tracing facilities such as the /proc file system provided by many Unix
operating systems. The behavior of some intercepted system calls is
then modified to implement new functionality.
This approach does not require any re-linking or
re-compilation of existing applications. In fact, the extensions
can even be dynamically ``installed'' into already running processes.
The extensions work completely at the user level and install without system
We used this approach to implement a global file system,
called Ufo, which allows users to treat remote files exactly as if they were
Currently, Ufo supports file access through the FTP and HTTP protocols
and allows new protocols to be plugged in. While
several other projects have implemented global file system abstractions,
they all require either changes to the operating system or modifications
to standard libraries.
The paper gives a detailed performance analysis of our approach to
extending the OS and establishes that Ufo introduces acceptable
overhead for common applications even though intercepting system
calls incurs a high cost.
/proc file system,
global file system,
global name space,
Computer users have always had the desire to extend operating systems
functionality to support new protocols or meet new usage patterns. In
this paper we show how to extend a standard
Unix operating system (Solaris) completely at the user level. Our approach --
which is similar to interposition agents [Jon93] -- uses tracing facilities to
intercept selected system calls at the user level. The behavior of
intercepted system calls is then modified to implement new
functionality. We use this to implement Ufo, a global file system,
which supports file access through
the FTP and HTTP protocols and can easily be expanded to support
While this paper focuses on extending the file system services, our
approach provides a general way of expanding operating system
functionality without any kernel changes or library modifications.
Our extensions, which are not just limited to Solaris,
do not require any re-linking or re-compilation of
existing applications. In fact, they can even be dynamically
``installed'' into already running processes. Extensions can be added
on a per-user basis, i.e., extensions for one user do not affect other
users. Actually, even a single user could run different jobs with
different extensions without interference.
An important advantage of this method is that developing the
OS extensions can be done entirely at the user level and without
access to OS source code.
This makes our approach an
excellent way for testing new kernel extensions and for providing
OS extensions that are not performance critical.
With the recent explosive growth of the Internet an increasing number
of users, including us, have access to multiple computers that
are geographically distributed. The initial motivation for our work was the
desire to have transparent file access from our Unix machines to our
personal accounts at remote sites. In addition, we also wanted to
present the resources from the large number of existing HTTP and anonymous FTP
servers as if they were local files. This would allow all local applications
to transparently access remote files.
Ufo implements a global file system that provides this
functionality. It is a user-level process that runs on multi-user Unix
systems and connects to remote machines via authenticated and
anonymous FTP and HTTP protocols. It provides read and write caching
with a weak cache consistency policy.
It was important to us that the file system not only run at the
user-level but that it also be user-installable
(i.e. installing it does not require root access).
For example, assume one of us obtains a new account at an NSF
supercomputer center. Once we log into that account,
we would like to transparently see all remote files we have some way of
accessing (be it via telnet, FTP, rlogin, NFS, or HTTP);
all without having to ask the system administrator to install
This is not necessarily easy to do in a Unix environment
since most current file system
software must be installed by a system administrator.
For example, systems such as NFS and AFS allow
sharing of files across the Internet, but they require root access to
mount or export new file partitions.
The system administrator
may not have the time or, due to security concerns,
may not be willing to install a new
piece of software or export a file system resource.
A user-installable file system does not have these problems.
Not only can users install it themselves, but it does
not introduce any additional security holes
in the underlying operating system or network protocol.
To guarantee that a file system can indeed be installed by the user,
it should only rely on functionality provided by standard (unmodified)
In order to provide a global file system we need extensions to the
operating system that handle file accesses (and related functions)
By modifying the behavior of the system calls, we can
add new functionality to the operating system. In our
approach we modify the system call behavior by inserting a user-level
layer, the Catcher, between the application and the operating system.
The Catcher is a user-level process which
attaches to an application and intercepts selected
system calls issued by the application.
From the user's perspective,
the Catcher provides
a user-level layer between the user's application processes and the
original operating system, as shown in Figure 1. This
extra layer does not change the existing OS, but allows us to
control the user's environment, either by modifying function
parameters, or issuing additional service requests.
A new view of the operating system.
The Catcher operates as follows:
Initially, it connects to the user process and tells the operating system
which system calls to intercept.
Our implementation which runs under Solaris 2.5.1 uses
the System V /proc interface, which was originally developed for debugging purposes [FG91].
Instead of just tracing the system calls, we actually change at the
user-level the semantics of some of them to implement the global file
Whenever a system call of interest begins (or completes),
the operating system stops the subject process and notifies the Catcher.
The Catcher calls the appropriate extension function, if needed,
and then resumes the system call.
For our global file system,
we intercept the open, close, stat, and
other system calls that operate on files.
When we intercept a system call which
accesses a remote file, we first ensure that an up-to-date
copy is available locally. Then, we patch the system call to refer to
the local copy and allow it to proceed. System calls which only
access local files are not modified (they can just continue), while
most systems calls not related to files are not even intercepted.
Since no application binaries are changed, this approach works
transparently with any existing executable (with the exception of the
few programs requiring setuid).
A potential concern with our approach is its performance overhead.
While the cost for intercepting system calls is significant,
our performance analysis shows that Ufo introduces acceptable
overhead for common applications.
runs fully at the user-level, if one user runs it there is no
performance penalty on another user. Furthermore, a user
can run Ufo only on some selected applications without impacting other
applications, or even dynamically ``install'' (attach) or ``uninstall''
(detach) Ufo while applications are running.
Installing Ufo can be done by any user
without needing root assistance.
The simplest way to start using Ufo is to explicitly
start processes under its control, e.g.
tcsh% ufo csh
csh% grep UCSB http://www.cs.ucsb.edu/index.html
csh% cd /email@example.com/
csh% emacs papers/ufo/introduction.tex &
In the example above, the new shell running under Ufo
can use the global file system's services. Ufo automatically
attaches to any child that the shell spawns,
like the grep and emacs processes above.
Alternatively, Ufo can be instructed to dynamically attach to
an already running process by providing its pid.
tcsh% emacs &
tcsh% ufo -pid 728
The remainder of the paper is structured as follows.
Section 2 reviews related work and compares
our method for operating system extension with
Section 3 describes in detail the Catcher
and how it intercepts system calls at the user-level.
Section 4 discusses the design decisions of Ufo, the
user-level global file system.
Section 5 presents experimental results
for a variety of micro-benchmarks, standard Unix file system
benchmarks, and full application programs.
Section 6 concludes this paper and
offers an outlook on future research directions.
Different methods of extending operating system functionality and examples.
Before presenting implementation details of the Catcher
(in Section 3), we
will put our work in context by comparing our approach with
alternative ways of extending operating system functionality.
The eager reader can skip directly to the discussion of
our implementation in Section 3.
We first introduce a classification of
different approaches to extending the operating system.
We then discuss the relevant research projects on
extending operating system and file system functionality
in more detail.
Different methods of extending operating system functionality and
There has been a considerable amount of work on extending operating
systems with new functionality. We can classify these approaches
into the following categories:
- Change Operating System:
The most straightforward approach is to
just modify the operating system itself and incorporate the
desired functionality. This requires access to the OS sources and the
privileges to install the new kernel.
- Device Driver:
- Instead of changing the kernel itself, modifications
can be limited to a new device driver which implements the desired
Root access is required to install the device drivers.
- Network Server:
A clean solution with minimal intrusion to the operating system is to
install a network server, which provides the additional services
through an already existing standardized interface.
Installing the server and
mounting remote directories requires root capabilities.
We want to re-iterate, these first three approaches require super-user
intervention and affect everybody using the system, since everybody
will see the modifications to the operating system. If there
is a bug or security hole in the newly installed software, the whole
system's integrity and security can be compromised.
A user-level approach avoids this problem.
- User-level Plug-Ins:
When a one-time modification to the operating system can be tolerated,
a flexible strategy is to add hooks to the operating system so that
system calls can trigger additional functions that extend the
This approach is especially appropriate if the OS has already been designed
to be flexible and support extensions.
- User-level Libraries (Static or Dynamic Linking):
Most applications do not directly access the operating system, but
use library functions embedded in standard
libraries. Instead of modifying all binaries or the OS kernel,
it suffices to make changes to the libraries.
Super-user privileges are only necessary if the original libraries/binaries
need to be replaced.
- Application Specific Modifications:
Instead of incorporating the modifications into the library, we can
also incorporate them directly into the application, avoiding
the operating system altogether.
- Intercept System Calls:
Most modern operating systems provide the functionality of
intercepting system calls at the user level.
A process can be notified
when another process enters or exits selected system calls. While the
original motivation for this functionality was debugging and
tracing of system calls, this mechanism can also be used to alter
their behavior. This mechanism, which first was used
in the context of Mach to implement interposition agents [Jon93],
forms the basis for our Ufo implementation.
Table 1 lists examples of the above approaches,
while Table 2 summarizes their limitations
and identifies the context in which they can be applied.
We wanted an approach which works with most existing applications
without the need for recompiling, and more importantly, which can be
used without requiring root access. Therefore we
decided to use the mechanism of intercepting system calls.
The project that is the closest to our own is the work on interposition
agents [Jon93] which also makes use of the mechanism of
intercepting system calls. Interposition agents provide a general
system call tracing toolbox, which allows different system calls
to be intercepted and handled in alternate ways, as we do in
Ufo. Three example agent applications were implemented: spoofing the
time of day, tracing system calls (as in truss), and transparently merging
the contents of separate directories.
The interposition agents work is based on Mach. While Mach is
a Unix variant, it was designed to be more flexible and extensible.
In particular, when calls are intercepted in Mach 2.5,
they can be redirected to the process' own address space. Thus,
the interposition agents are run in the user process' own memory.
Approaches that use a more standard Unix, such as ours, are more constrained
(and more complicated to implement) since it is more difficult to access
the user process state from outside of the process' address space.
Another research project that uses
the Unix trace mechanism for implementing an OS extension is Janus
[GWTB96] which provides a secure, confined environment for
running untrusted applications safely by intercepting and selectively
denying system calls. Like ours, the Janus implementation has been
designed for Solaris.
A lot of current
research deals with designing operating systems such that they allow for easier
and more efficient user-level extension. Engler et al. [EKO95]
carry the Mach micro-kernel methodology
[ABB 86] further by removing as much kernel
abstraction as possible from the OS. This pushes the kernel/user-level boundary
as low as possible, placing most of the OS services outside of the kernel.
Another approach, taken by VINO [SESS94]
and SPIN [BSP 94], is to allow injection
of user-written kernel extensions into the kernel domain. A discussion
of the issues involved can be found in [SS96].
Another recent project, SLIC [GPA96], is an
OS extension to Solaris
that allows for plug-ins at both the user and the kernel level.
We now discuss operating system extensions
specific to our particular application: remote file transfer.
There are a number of systems that provide
transparent access to remote resources on the Internet, many of
which have been very successful. Examples
include NFS [SGK 85], AFS [MSC 86],
Coda [SKK 90], ftpFS in Plan 9
[PPTT90] and Linux [Fit96], Sprite
[Wel91, NWO88], WebFS [VDA96], Alex
[Cat92], Prospero [NAU93], and Jade [RP93].
They all have one significant drawback, however:
they either require root access or modifications
to the existing operating system, applications or libraries.
Ufo is distinct in that it requires no such modifications
to any existing code and runs entirely at the user-level.
There are a few systems for global file access
that run entirely at the user-level and are user-installable.
They are also similar to Ufo in that they
extend a local file system to provide uniform and
transparent access to heterogeneous remote file servers.
Prospero [NAU93] and Jade [RP93]
both provide access to NFS and AFS file systems, and to FTP servers.
Prospero runs at user-level by replacing standard statically linked libraries.
This avoids changes to the operating system, but
requires re-linking of existing binaries.
Jade [RP93] uses dynamic libraries instead
and allows most dynamically linked binaries to run unmodified.
Changing application libraries works well for
most applications, especially when combined with dynamic linking.
The drawback of this approach is that it does not work
for statically linked applications not owned by the user as well as
for applications that circumvent the standard librtaries and
execute system call instructions directly.
Other global file systems also run at the user level, but are not
user-installable, since they require extensions to the operating system itself,
which in turn requires root access.
One such example is
WebFS [VDA96], a global user-level file system based on
the HTTP protocol. To run at the user level, WebFS relies on the OS
extensions provided by SLIC [GPA96], which implements a
call-back mechanism to a user process.
(WebFS also requires the HTTP server
be extended with a set of CGI scripts that service requests.)
Similar to SLIC, UserFS [Fit96]
is an OS extension that enables user-level file systems to
be written for Linux.
While installing UserFS itself requires kernel recompilation,
installing new file modules, such as ftpFS, does not.
Plan 9 [PPTT90] also includes an FTP based file system
(also called ftpFS). At least two projects provide access to FTP servers by implementing
an NFS server that functions as an FTP-to-NFS gateway.
Alex [Cat92] supports read-only access to
anonymous FTP servers, while [Gsc94] additionally
allows read and write access to authenticated FTP servers.
In this section we discuss the details of our implementation of the Catcher
inside Ufo. We start by describing the high-level architecture
and the role of the Catcher in Ufo.
Ufo is a user-level process which provides file system services to
other user-level processes by attaching to them. Once
attached to a subject process, it intercepts system calls and
services them if they operate on remote files. The application is
unaware of the existence of the Ufo, but, with Ufo's help, it can operate on
remote files as if they were local.
General architecture of Ufo.
Ufo is implemented in two modules: the Catcher and the Ufo module
(Figure 2). The Catcher is responsible
for intercepting system calls and forwarding them to the Ufo module.
The Ufo module implements the remote file system and
consists of three layers: the File Services layer
which identifies remote files, the Caching layer, and the Protocol layer
containing different plug-in modules implementing the actual file
Figure 2 shows the steps involved in servicing
a remote file request. When the application issues a system call (1),
it can go directly to the kernel or, if it is file-related, get
intercepted by the Catcher (2). For intercepted calls, Ufo determines
whether the system call operates on a remote or a local file, possibly
using kernel services (3,4). If the file is local, the request proceeds
unmodified. If the file is remote, Ufo creates a local
cached copy, patches the system call
by modifying its parameters, and lets the request proceed to the kernel
(5). After the
request is serviced in the kernel (6), the result is returned to the
application (7). The return from the system call may also be intercepted and
patched by Ufo, though the figure does not show this.
In our Solaris implementation, the Catcher monitors user processes using
the /proc virtual file system [FG91].
This is the same method used by monitoring
programs, such as truss or strace, which are also available on a number
of other UNIX platforms, including Digital Unix, IRIX, BSD or Linux.
The System V /proc interface allows us to monitor and modify
an individual process by operating on the file associated with a
In particular the Catcher attaches to
a subject process pid by opening the /proc/pid file.
Once attached, the Catcher uses ioctl system calls
on the open file descriptor to control the process. It can
instruct the operating system to stop the subject process on
a variety of events of interest. In Ufo there are two events of
interest: system call entry into the kernel, and
system call exit from the kernel.
Once a subject process has stopped on an event of interest,
the Catcher can read and write the registers and read and write
in the address space of the process.
The Catcher uses this to examine and modify the parameters or the result of
system calls like open, stat, and getdents.
Finally, the /proc interface allows us to restart
the execution of a stopped process. Figure 3
summarizes how the discussed functionality is used in the
Outline of the Catcher algorithm.
Conceptually, Ufo implements the system calls intercepted
by the Catcher, but in practice Ufo does not service them directly.
Although implementing the system calls directly in Ufo would be possible,
it would require reimplementing existing OS functionality.
Instead, we ``patch'' the system calls by (i) modifying the
call's parameters, (ii) changing the file system state
(e.g., fetching a file from a remote server)
and (iii) modifying the result returned from the operating
A good example for the first two actions is the open system
call. On the entry of an open call, we may have to
modify the file name string to point to the locally cached copy.
Before allowing the system call to continue, the Catcher may have
to wait for Ufo to download the file from the remote site.
Implementing the name change is somewhat complicated since we must modify the
user's address space. We cannot just change the filename in place since
the new filename might be longer than the old one. Also, the filename
could be in a segment which is read-only or shared among threads.
Currently we solve these problems by writing the new file name
in the unused portion of the application's stack and changing the
system call argument to point to the new string.
The open system call needs to be intercepted on exit from the
kernel as well. Although the returned result
is not modified, Ufo must remember the correspondence
between the returned file handle and the file name, which is
needed when the file is closed.
Besides file related system calls, there are several others that
must be intercepted. For example, to track child processes
we intercept the fork system call. Given the
child pid, we can open its associated /proc file and monitor it as
well. System V allows the set of trapped system calls to be automatically inherited
from parent to child, so this setup is only needed for the initial
Implementations of file system functionality at the user level
must obey some restrictions since user-level processes cannot
perform arbitrary actions and cannot access the whole file system
related state of the kernel.
One problem is that the Catcher cannot control setuid process since
the security policy of the operating system disallows user-level
processes from attaching to other users' processes. In practice we
have found this not to be a problem since very few programs are
installed with setuid. And for most of those programs, e.g., rlogin,
it is not clear whether one really needs the file system extensions. In
the current implementation, whenever the Catcher detects that a
subject process is about to spawn a setuid program, it just does
not trace the child process.
In Solaris, the /proc interface allows the controlling process to
write in the subject process' address space. This is important if the
Catcher needs to change some system call arguments such as filename
strings for Ufo.
If we are to port the Catcher to other operating systems that
do not provide the capability of writing into the subject process,
we would not be able to implement this feature.
Although writing in the user process is not necessary for
the basic functionality of Ufo, a Ufo implementation on such
an operating system would have some limitations. Features such as the
URL naming scheme and mountpoints in the root directory
require changing of string arguments to system calls and therefore
would not be possible to implement (see Subsection 4.1).
Another problem arises when the Catcher process is killed.
Being a regular user-level process, the Catcher
cannot protect itself against the SIGKILL signal. There is no graceful
way to handle such a situation if the subject processes running under
the Catcher continue working on remote files. In the current Ufo
implementation the subject process will be trapped on the next
intercepted system call and stay trapped until killed.
The Catcher mechanism allows us to create a personalized operating
system. Requests made to the kernel can be re-interpreted, in effect
allowing individual users to run their own OS. Any user can use the
``new'' OS without having to modify the original operating system or
needing root access. Although, the current Catcher only intercepts
system calls, System V allows the user to also intercept and act on
signals and hardware faults. This allows for a wide range of
OS functionality to be extended using the Catcher mechanism.
Other potential uses of the Catcher for personalized OS extensions include
encrypting file systems, file systems which store files in
compressed form, confined execution environments for runing
untrusted binaries [GWTB96],
virtual memory paging [DWAP94, FMP 95], and
process migration [Con95].
A potential concern with our approach is its performance overhead.
Indeed, intercepting individual system calls is quite expensive and
for some OS extensions this overhead would be unacceptable.
Nevertheless, Ufo is an example that there are OS extensions for
which the Catcher mechanism works well. Our
performance analysis shows that Ufo introduces moderate
overhead for common applications. This is due to
the fact that typical applications issue relatively few
system calls, and not all system calls are
intercepted in Ufo.
provides read and write access to FTP servers and read-only access
to HTTP servers. The remote file access functionality is implemented
in Ufo's file system module which is responsible for resolving remote
file names, transfering files, and caching.
Ufo supports three ways of specifying names of remote files:
(i) through a URL,
(ii) through a regular filename implicitly containing the remote host,
user name, and access mode, and (iii) through mount points.
The first way to specify a remote file is through its URL syntax.
Unfortunately, some applications cannot handle URL names. Make
and gmake cannot handle the colon in the URL, while Emacs considers // to be
the root of the file system and thus discards everything to the left.
To alleviate these problems we also support specifying a remote file
through a regular filename. The general syntax is
where protocol is the file transfer protocol,
e.g., ftp or http.
Lastly, Ufo allows the user to specify
explicit mountpoints for remote servers or access protocols
in a .uforc file.
For example, the line
local /csftp remote /
machine ftp.cs.ucsb.edu method FTP
specifies that accesses relative to /csftp refer to
the root directory of the ftp.cs.ucsb.edu
anonymous FTP server. The user
can also specify mountpoints for access methods. In fact
that is how the second naming scheme is implemented:
if the user does not explicitly specify a mount point
for the HTTP method, for example, Ufo uses the
local /http method HTTP
Similarly to Sprite [NWO88], we have implemented mount points using
a prefix table which, given a filename, searches for the longest
matching prefix in the list of mount points.
Ufo also supports symbolic links. A user
can create links to frequently accessed remote directories. While links
simplify accesses to remote files, they actually present quite an
implementation challenge, since they require following all link
components to determine the true name of a file.
Ufo transfers only whole files to and from the remote file system.
Whenever Ufo intercepts the open system call
for a remote file, it ensures that
a local copy of the file exists in the cache,
and then redirects the system call
to the local copy. Read and write
system calls don't even have to be
intercepted since they operate on file descriptors returned by
they will correctly access the local copy in the cache.
Finally, on a close system call, Ufo checks whether the file has been
modified and if so, stores the file back to the server
(the store may be delayed if write-back caching is in effect).
Ufo uses whole file transfers for two reasons:
this minimizes the number of system calls that need to be intercepted,
and protocols such as FTP only support whole file transfers.
When an application requests information about a remote file, e.g., through
a stat or lstat system call, Ufo satisfies the
request by creating a local file stub and redirecting the system call
to it. The file stub has the correct modification date and size of the
remote file but contains no actual data.
With this approach Ufo neither has to re-implement the stat system
call, nor download the whole file. Only if the application wants to
open a file stub later, will Ufo actually download the remote file.
Similarly, when a system call such as getdents (get directory entries)
is issued on a remote directory, Ufo creates a copy of the directory
in the local cache and puts file stubs in it. Then, it redirects the system
call to the so created skeleton directory.
Since remote data transfers can be quite slow,
Ufo implements caching of remote files to achieve reasonable
Instead of downloading a file each time the user
opens it for reading, Ufo keeps local copies of previously accessed files.
Ufo can reuse the local copy on a subsequent access,
as long as it is up-to-date.
Similarly, we use write-back caching which delays writing
a modified file back to the remote server.
While files are the primary objects cached, Ufo also caches
directory information (directory contents), and file information
(size, modification time, permissions).
The FTP module additionally caches open control connections.
Since establishing a new connection to
the remote server for each transfer is expensive, we
reuse open control connections by keeping them alive
for a period of time after a transfer has completed.
The cache consistency policy governs whether we are allowed to use a
local copy on a read, and whether we can delay the write-back of a
To efficiently support a wide range of usage patterns, Ufo provides an
adjustable consistency policy based on timeouts (a read and write delay).
The policy guarantees that
(i) when a file is opened it is no more than
seconds out of date; and
(ii) changes made to a file will be written back to the server
within seconds after the file is closed.
To verify that a local file
is up to date (i.e., is not stale), Ufo checks whether the file on
the remote site has changed (validate on open).
and can have a zero value. In this case files
opened for reading are never stale and modified files are written back to
the server immediately after they are closed.
The write timeout of a file
is always a certain number of seconds.
The read timeout can optionally
be specified as a percentage of the file's age as in Alex
[Cat92]. This method is based on the observation that older files are
less likely to change than newer files. Therefore older files
need to be validated less often.
Files can have individual timeouts and
Ufo provides mechanisms for the user to define default timeouts for all files,
or for all files on a server.
This allows the user to adjust the tradeoff between performance
and consistency based on known usage patterns. For example,
when mounting read-only binaries
large read timeouts can be used since
these files change rarely.
Ufo relies on the underlying access protocols for authentication.
Currently, passwords are only required for authenticated
FTP servers and are not needed for HTTP and anonymous FTP
accesses. Ufo allows the passwords to be stored in the
.uforc or .netrc files,
or alternatively, Ufo asks for the password on the first
access to a remote server.
Since Ufo is running entirely at the user level with the access
permissions of its owner, it does not introduce new security problems
in the system. The only potential security concern
is to ensure that other users do not gain undesired access to the files
in the private Ufo cache. We avoid this problem
by creating the topmost cache directory with read and write
permissions for the owner only.
In implementing Ufo, we tried to minimize the amount of
operating system functionality that we had to reimplement. First,
we attempted to minimize the number of intercepted system calls
in order to minimize the execution overhead
that Ufo introduces. This
lead to the whole file caching policy.
Second, we wanted to minimize the implementation effort
by modifying/reimplementing as few system calls as possible.
This lead to our decision to create file stubs and skeleton directories
for the stat and getdents calls.
Of course, there is a trade-off between execution overhead
and implementation effort.
For example, the advantage of creating file stubs and skeleton directories is
that we do not have to reimplement the stat and getdents
system calls. The disadvantages are that
creating file stubs may have high overhead.
Also for efficiency, we rely on the support for
holey files by the local file system.
For example, on our machines the /tmp file system
does not support
holey files, thus if we use /tmp for the Ufo cache the stubs for
large files do use all the disk space indicated by their size. The
NFS-mounted file systems at our site do support holey files, but
the stub creation there is an order of magnitude slower than on /tmp.
For these reasons we are considering implementing the
stat and getdents system calls completely inside
the next version of Ufo to improve its performance. In fact, we are already
partially implementing (patching) the getdents system call
in order to support Ufo mountpoints in user-unwritable
areas such as the root directory.
Transferring only whole files
introduces three well known problems for extremely
large files [Cat92].
First, when only a small fraction of a file is actually
accessed, a lot of unnecessary data may be transferred.
Second, the whole file has to fit on the local disk.
In practice we don't expect these two problems to occur frequently.
With the exception of databases, most
applications tend to access files nearly in their entirety [BHK 91].
Furthermore, Ufo allows any local file system to be
used for file transfers, thus reducing the danger of insufficient local
A third problem comes from our decision not to
intercept the read and write system calls.
In our approach the open call
blocks until the whole file has been transferred.
It is possible to intercept and handle
read and write system calls in Ufo. The benefit
is that open would not
reads that operate on the already present part of a file
could be executed without waiting for the completion of the whole transfer
(see Alex [Cat92]). The drawback is that intercepting
read and write calls incur a high overhead and
requires extra implementation effort.
Run times in microseconds for various system calls for accessing files in /tmp
(the numbers are the arithmetic mean of 5 runs, each executing 100 iterations).
The numbers in parentheses represent the ratio normalized
to the standard Solaris OS.
The main goal of our performance analysis is to measure the
overhead introduced by the Catcher mechanism in Ufo. This information
is necessary to determine the usability of our method for
operating system extension.
We first present the results of several microbenchmarks,
which measure the overhead of intercepting individual Unix system calls.
To demonstrate the overall impact of this overhead on whole applications
we also present measurements for a set of file system benchmarks and a set
of real-life applications. While the microbenchmarks show that
intercepting system calls is expensive, the real-life applications
exhibit much lower overhead.
All tests were run on a 143 MHz Sun Ultra 1 workstation with 64 megabytes
of main memory running Solaris 2.5.1.
The microbenchmark results present the user-perceived run times
(measured as wall clock times) for
open, close, stat, read, write, and
getpid system calls. The
results are shown in Table 3.
show the numbers for the normal user program, for the
Catcher-monitored program (Catcher only, no calls to Ufo functions),
and for the Ufo program (Catcher and Ufo functionality). In the latter
case, we examine the run times for a local file, for a cached
remote file and for a remote file that has not been cached.
The Catcher only and Ufo local file numbers are of
special significance. They show the cost of running
a process under the Catcher or under Ufo when the process
accesses local files only and does not require
any of the extended OS functionality. This is the
the fundamental overhead introduced by our
method of extending the OS.
The numbers for remote files are a measure
of the combined effect of our remote file system implementation,
our caching policy, the efficiency of the underlying access
protocol (FTP in this case), and the quality of the network connection.
In order to measure the cost of the Solaris system calls themselves and
not the network speed or the NFS overhead, we used the local
/tmp file system. Accesses to /tmp are very fast and
do not involve disk, network traffic or protocol overhead.
As a result the microbenchmarks present the Catcher and Ufo overhead
in the worst-case scenario. The relative Catcher and Ufo overhead for accessing
non-cached NFS files, for example, is much lower.
The microbenchmarks were run on a lightly loaded workstation
by taking the wall-clock time just before and just after the system call.
The timing was done using the high resolution timer
gethrtime which has a resolution of about
0.5 microseconds on the Ultra 1 workstation.
Since individual system calls are very fast, normal system activity
such as interrupts and context switches distorts some of
the measurements. This produces a small percentage of outliers that
are several times larger than the rest of the measurements. To
ensure we do not include unrelated system activity in our
measurements, in each test run we recorded 100 measurements
and discarded the highest 10% of them. The remaining times were then averaged.
The numbers in the table are the arithmetic mean of five such runs. The standard
deviation for the five runs was below 2% for all tests, except for
getpid, for which the standard deviation was at most 6%.
The Catcher only numbers show the cost of intercepting system
calls. The results are obtained by running the benchmark program under
the control of the Catcher alone. The Catcher simply intercepts the
open, close and stat system calls executed by the benchmark program,
and lets them continue immediately without modifying them.
The read, write and getpid system calls are not intercepted at all.
Even though one may expect that these system calls will not be affected,
they do incur a small overhead: whenever there is
even a single intercepted system call for a process, the operating
system takes a different execution path for all system calls
of that process, independent of whether they are intercepted or not.
The results demonstrate that for read and
write of 1 byte blocks this overhead is small and for
8K blocks it is negligible.
Because getpid is so fast, it has a substantial
relative overhead, but still only 2 total.
On the other hand, system calls that must be trapped by the
Catcher incur a factor of 4-9 overhead.
During this extra time, control is passed from the program to the Catcher,
(which performs ioctl calls to read information from the /proc
file system), and then back again.
The Ufo local file column shows how much extra overhead is introduced
by Ufo in addition to the Catcher. The benchmark program is running
under Ufo and is accessing local files only. Even though no
remote files are accessed, Ufo still introduces some overhead in addition
to the Catcher overhead. The extra overhead comes from the
analysis of the parameters of the intercepted system calls.
For system calls that reference a file, Ufo determines
whether the file is indeed local or remote. Since a system call
does not necessarily take an absolute path name as an argument, Ufo
has the responsibility of determining it.
Determining the true filename can involve a number of stat
system calls, similar in flavor to the pwd command, and this can
add a noticeable overhead.
The remaining two columns measure the overhead of Ufo when working
with remote files.
These numbers are measured as with Ufo local file, except that
the accesses are to remote files. For the
Ufo remote cached tests, a locally cached copy of the remote file is
Note that in either case (cached or uncached),
the read and write system calls operate on the locally cached copy
of the file. Thus, these numbers are consistent across all of the
tests. On the other hand, open and stat
calls to uncached remote files require remote accesses, and the overhead
increases dramatically when Ufo uses the FTP protocol
to retrieve the file.
This overhead is almost entirely determined by the quality of the
network connection and the FTP protocol.
In our measurements we accessed files located at UC Berkeley.
From a UC Santa Barbara machine, opening a remote file of size
1024 bytes residing at a UC Berkeley host requires 531ms using FTP.
Closing the same remote file after modifying it takes 452ms
since the file must be written back to the remote server.
If the file is cached, the open, close and stat overhead
is much smaller, but it still has roughly four times the overhead compared
to a local file. This is due to two reasons: the additional work to manage the
cache, and several remaining inefficiencies in our prototype implementation which
will be corrected in future versions of Ufo.
Run times for the Iostone and Andrew file system benchmark programs with and without
Ufo. Times are in seconds, with the ratios normalized to the standard
OS shown in parenthesis. ( The Andrew benchmark reports its timing results with
a resolution of 1 second. The 0 seconds in the table indicate a measurement between 0 and
Table 4 reports the absolute execution times in
seconds for two file system benchmarks run on the local /tmp file
system with and without Ufo and on a remote FTP-mounted file system
with and without caching. For these tests, the FTP host
was a machine on the local 100Mbit/s Ethernet network.
The remote tests with caching were with
a warm cache and read and write delays set to infinity. Thus, these
measurements represent the best-case scenario for remote files. For the
remote tests without caching, the read and write delays were set to
zero, forcing every open, close and stat system calls
to go to the remote site. These tests are the worst-case scenario for
accessing remote files under Ufo.
Iostone and Andrew are standard file
system benchmarks. We chose these as examples of applications that
execute a lot of file system calls that Ufo intercepts and handles.
The Iostone benchmark [IOS87] performs thousands of file accesses (opening,
reading, and writing). Because of the large amount of file opens
and closes, Ufo runs about 8 times slower on the local file system.
The Andrew benchmark [HKM 88]
measures five stages in the generation of a software tree.
The stages (i) create the directory tree, (ii) copy source code into the
tree, (iii) scan all the files in the tree, (iv) read all of the files,
and finally (v) compile the source code into a number of libraries.
For this benchmark the Ufo overhead on local files is a factor of 1.33, much
lower than the overhead for Iostone. For both Andrew and Iostone,
the results for the uncached remote tests are orders of magnitude worse than for
the local /tmp file system.
This is not
not surprising since the network latency and the FTP protocol overhead are
quite large compared to the fast accesses in /tmp.
We also tested Ufo with a number of larger Unix applications:
latex, ghostscript, a make of the Ufo
executable, and the integer applications from the SPEC95 benchmark suite.
The results are shown in Table 5.
As with the file system benchmarks,
each test was run without Ufo, under Ufo on local files only, and
under Ufo on remote files with and without caching.
The first set of benchmarks are programs that we run frequently.
The latex test measures the time to latex three times a 20 page paper consisting
of 8 tex files and
then produce a postscript from the dvi file.
The make test compiles Ufo itself using g++. The ghostscript
test displays a 20 page postscript document.
The table shows that latex and make perform a relatively
large number of system calls that Ufo intercepts,
mainly open, close, and stat.
This results in Ufo overheads of 24% and 22% respectively,
when run locally, and higher overheds, when run remotely.
The remote overheads, while large,
should be acceptable to the user, since accessing remote files
is expected to cost extra time. The local overheads on the
other hand, are incurred only because the application is
running under Ufo even though it is not using any of its functionality.
To avoid unnecessary local overhead,
applications that only access local files can be run without Ufo,
and Ufo can be detached from applications once they stop
accessing remote files.
The ghostscript test on the other hand performs few
calls that Ufo intercepts and never writes to the remote
server; as a result the Ufo overhead is
very low even in the remote test. This sort of overhead should
be unnoticeable to the user.
The last eight tests are the integer applications from the SPEC95
benchmark suite. These were chosen as examples of
compute intensive applications that
do not perform extensive file system operations. For these applications
the observed overhead is very small in the local and even in the
Small perceived overheads should also be expected for interactive
applications such as text editors since
the user is not likely to notice the difference between
28 and 611 when opening a local file.
Relative run times for some file system benchmarks and larger Unix applications. Times are in
seconds, and the relative speed in parentheses. The first column shows
the number of system calls executed by the application.
As expected, we find that intercepting system calls
can be very expensive, and remote accesses are orders of magnitude higher
than local accesses.
For programs such as the Iostone benchmark -- which performs many
open, close and stat calls -- the Ufo overhead
for local files is too large to be ignored.
Clearly, such programs should not be run under Ufo if they only access
local files since this will incur a large overhead even though the program
does not utilize any of the extended functionality.
If remote files need to be accessed, then programs like Iostone will
run slow, but this is mainly due to the network latency and access protocol
overhead which by far outweighs the Catcher and Ufo overhead as shown in
Table 3. In this case Ufo proves to be
a convenient tool.
Furthermore, Ufo allows the user to dynamically attach to
a running process and detach from it,
so the choice of running under Ufo or not
is always available. Applications that need access to remote
files can have it, and the remaining processes will not incur
Other applications, such as make, and latex incur a 22-24%
overhead on local files -- noticeable, but perhaps acceptable to the user
even when the functionality of Ufo is not required. For remote files
these applications incur overheads of 24% in the best case and 600% in the
worst case depending on the kind of file caching used. In most
cases the user expects that working on remote files would be slower, so
the use of the extra functionality provided by Ufo
should be worth the additional overhead, especially when
the only alternative is to manually transfer files using FTP.
Many other applications, such as compute intensive programs or
text editors, make infrequent use of the system
calls trapped by Ufo (though they may use other calls
such as read and write). For such applications,
user-perceived delays are much smaller: on the order of a few percent.
In this case, running applications under Ufo makes no appreciable
From these observations we can draw the conclusion that
the Catcher is a good tool for implementing operating system
extensions that require the interception only of relatively
infrequent system calls.
An example of such an extension
is the Ufo file system when running real-life applications.
On the other hand, this method is not ideal for
extensions which intercept frequently occurring system
Finally, we would like to mention that we are aware
of some opportunities for improving the
current Catcher prototype and many opportunities for improving
Ufo. For example, optimizing the filename check in Ufo
(to determine whether a file is local or remote) alone will
result in a significant reduction of the running time.
In this paper, we presented a general way of extending operating
systems functionality, using the debugging and tracing facilities provided by many Unix
operating systems. Selected system calls are intercepted at the
user-level and augmented to obtain the desired functionality.
This mechanism forms the basis for Ufo, a file system providing transparent
access to remote files on FTP and HTTP servers. Ufo proved to be a useful
tool which we now use daily.
As our experimental results show, its overhead, while quite
large for intercepted system calls, is acceptable for most applications.
We believe that our approach is a promising way for individual users
to develop and experiment with future operating system extensions, since
this can be done completely at the user-level. Essentially, each
user sees a personalized version of the operating system,
extensions do not affect other users and are compatible with existing
applications as those need not be re-compiled or re-linked. In the
past, operating systems research had a hard time to carry over to the
general public. With our approach, researchers can make
their extensions easily
available, and users can run them without relying on the system
administrator for installation.
There are plenty of avenues for future work and research.
For example, we have several ideas on how to
improve the performance of the Catcher and Ufo.
We also plan to implement new protocol modules in
Ufo, e.g. based on NFS, WebNFS, and the rlogin protocols.
We have experimented with several other OS extensions suitable for
cluster of workstation environments. For example, we have developed a
prototype that attaches to a process, checkpoints it, and then can
restart it at a later time or migrate it to another processor.
Similiarly, we have a prototype Catcher which intercepts all forks and
execs and sometimes decides to execute some processes on other
workstations. While both tools are still at a very crude stage, we
have already seen some of their potential benefits.
Similar benefits can be expected for paging virtual
memory to the memory of idle processors instead of to a slow local disk.
Another interesting research area is protected computing. The system
calls define the capabilities a process has and resources it can
obtain (memory, disk access, CPU time). We can use the Catcher to
limit the resources a process can access or obtain. This
approach, implemented in Janus [GWTB96],
is especially interesting in the current
development of global computing, where one user may run an untrusted
binary fetched from the Internet.
Finally, we intend to generalize our design of the Catcher since it
can not only intercept system calls, but also
signals and hardware traps which are delivered to the application.
We intend to build a Catcher toolbox which can be used for OS
courses and research projects.
We would like to thank Urs Hölzle for insightful discussion
of this paper and for helping test early versions of Ufo. We would also
like to thank Dave Probert, Chad Yoshikawa, Roger Faulkner,
and the anonymous referees for their valuable feedback.
- ABB 86
M. Acetta, R. Baron, W. Bolowsky, D. Golub, R. Rashid, A. Tevanian, and
Mach: A new kernel foundation for Unix development.
In Proceedings of the USENIX Summer '86 Conference, July 1986.
- BHK 91
M. G. Baker, J. H. Hartmann, M.D. Kupfer, K. W. Shirrif, and J. K. Ousterhout.
Measurement of a distributed file system.
In Proceedings of the 13th Symposium on Operating System
Principles, Pacific Grove, CA, 1991.
D. R. Brownbridge, L. F. Marshall, and B. Randell.
The Newcastle Connection, or UNIXes of the world unite!
Software -- Practice and Experience, 12, 1982.
- BSP 94
B. N. Bershad, S. Savage, P. Pardyak, E. F. Sirer, M. E. Fiuczynski, D. Becker,
C. Chambers, and S. Eggers.
Extensibility, safety and performance in the SPIN operating
In Proceedings of the 15th Symposium on Operating System
Alex -- a global filesystem.
In Proceedings of the 1992 USENIX File System Workshop, Ann
Arbor, MI, May 1992.
The Condor Team.
Checkpoint & migration of UNIX processes in the Condor
distributed processing system.
Dr. Dobbs Journal, February 1995.
M. Dahlin, R. Wang, T. Anderson, and D. Patterson.
Cooperative caching: Using remote client memory to improve file
In Proceedings of the USENIX Conference on Operating System
Design and Implementation, May 1994.
D. Engler, F. Kaashoek, and J. O'Toole.
Exokernel: An operating system architecture for application-level
In Proceedings of the 15th ACM Symposium on Operating System
Principles, December 1995.
P. R. Eggert and D. S. Parker.
File systems in user space.
In Proceedings of the Usenix Winter 1993 Technical Conference,
Berkeley, CA, 1993. Usenix Association.
R. Faulkner and R. Gomes.
The process file system and process model in UNIX system V.
In Proceedings of the 1991 USENIX Winter Conference, 1991.
Userfs: A user file system for Linux.
- FMP 95
M. J. Feeley, W. E. Morgan, F. H. Pighin, A. R. Karlin, and H. M. Levy.
Implementing Global Memory Management in a Workstation Cluster.
In Proceedings of the 15th ACM Symposium on Operating Systems
Principles, December 1995.
D. P. Ghormley, D. Petrou, and T. E. Anderson.
SLIC: Secure loadable interposition code.
Technical Report CSD-96-920, University of California, Berkeley,
FTP -- access as a user-defined file system.
ACM Operating Systems Review, 1994.
I. Goldberg, D. Wagner, R. Thomas, and E. A. Brewer.
A Secure Environment for Untrusted Helper Applications --
Confining the Wily Hacker.
In Proceedings of the 1996 USENIX Security Symposium, 1996.
- HKM 88
J. H. Howard, M. L. Kazar, S. G. Menees, D. A. Nichols, M. Satyanarayanan,
R. N. Sidebotham, and M. J. West.
Scale and performance in a distributed file system.
ACM Transactions on Computer Systems, 6(1), February 1988.
A synthetic file system performance benchmark.
Technical Report TR-074-87, Princeton University, 1987.
M. B. Jones.
Interposition agents: Transparently interposing user code at the
In Proceedings of the 14th Symposium on Operating Systems
Principles, New York, NY, 1993.
- MSC 86
J. Morris, M. Satyanarayananan, M. H. Conner, J. H. Howard, D. S. Rosenthal,
and F. D. Smith.
Andrew: A distributed personal computing environment.
Communications of the ACM, 29(3), 1986.
B. C. Neumann, S. S. Augart, and S. Upasani.
Using Prospero to support integrated location-independent
In Proceedings of the Symposium on Mobile and
Location-Independent Computing, Cambridge, MA, 1993.
Free Software Foundation, Inc.
M. Nelson, B. Welch, and J. Ousterhout.
Caching in the Sprite network file system.
ACM Transactions on Computer Systems, 6(1), February 1988.
R. Pike, D. Presotto, K. Thompson, and H. Trickey.
Plan 9 from Bell labs.
In Proceedings of the UKUUG Conference, July 1990.
H. C. Rao and L. L. Peterson.
Accessing files in an internet: The JADE file system.
IEEE Transactions on Software Engineering, 19(6), June 1993.
M. Seltzer, Y. Endo, C. Small, and K. Smith.
An introduction to the VINO architecture.
Technical Report TR34-94, Harvard University, 1994.
- SGK 85
R. Sandberg, D. Goldberg, S. Kleiman, D. Walsh, and B. Lyon.
Design and implementation of the Sun network file system.
In Proceedings of the Summer USENIX conference, June 1985.
- SKK 90
M. Satyanarayananan, J. J. Kistler, P. Kumar, M. E. Okasaki, E. H. Siegel, and
D. C. Steere.
Coda: A highly available file system for a distributed workstation.
IEEE Transactions on Computers, 39(4), 1990.
M. Seltzer and C. Small.
A comparison of OS extension technologies.
In Proceedings of the 1996 Usenix Technical Conference, San
Diego, CA, January 1996.
A. Vahdat, M. Dahlin, and T. Anderson.
Turning the web into a computer.
Technical report, University of California, Berkeley, 1996.
B. B. Welch.
Measured performance of caching in the Sprite network file
Computer Systems, 3(4), 1991.
Extending the Operating System at the User Level:
the Ufo Global File System
This document was generated using the LaTeX2HTML translator Version 96.1 (Feb 5, 1996) Copyright © 1993, 1994, 1995, 1996, Nikos Drakos, Computer Based Learning Unit, University of Leeds.
The command line arguments were:
latex2html -split 0 -show_section_numbers top_html.
The translation was initiated by Albert D. Alexandrov on Thu Nov 14 13:29:37 PST 1996
- This work was supported by the National Science Foundation
under NSF CAREER Award CCR-9502661 and NSF Postdoctoral Award ASC-9504291.
Computational resources were provided
by the NSF Instrumentation Grant CDA-9529418 and
The software is available on-line under http://www.cs.ucsb.edu/research/ufo
- The acronym Ufo stands for User-level File
- Similar functionality is
provided by Digital Unix, IRIX, BSD and Linux. This mechanism is used by
system call tracing applications such as truss or strace.
a stub is done by seeking to the desired position in a newly created file
and then writing a single byte. On most file systems,
the so created stub occupies a small
amount of disk space, independent of the reported file size.
Several other system calls like
lseek also have to be intercepted for this to work.
Albert D. Alexandrov
Thu Nov 14 13:29:37 PST 1996