Home - Topics - Papers - Theses - Blog - CV - Photos - Funny

VXA: A Virtual Architecture for Durable Compressed Archives

Bryan Ford

Computer Science and Artificial Intelligence Laboratory
Massachusetts Institute of Technology

Abstract:

Data compression algorithms change frequently, and obsolete decoders do not always run on new hardware and operating systems, threatening the long-term usability of content archived using those algorithms. Re-encoding content into new formats is cumbersome, and highly undesirable when lossy compression is involved. Processor architectures, in contrast, have remained comparatively stable over recent decades. VXA, an archival storage system designed around this observation, archives executable decoders along with the encoded content it stores. VXA decoders run in a specialized virtual machine that implements an OS-independent execution environment based on the standard x86 architecture. The VXA virtual machine strictly limits access to host system services, making decoders safe to run even if an archive contains malicious code. VXA's adoption of a ``native'' processor architecture instead of type-safe language technology allows reuse of existing ``hand-optimized'' decoders in C and assembly language, and permits decoders access to performance-enhancing architecture features such as vector processing instructions. The performance cost of VXA's virtualization is typically less than 15% compared with the same decoders running natively. The storage cost of archived decoders, typically 30-130KB each, can be amortized across many archived files sharing the same compression method.


1 Introduction

Data compression techniques have evolved rapidly throughout the history of personal computing. Figure 1 shows a timeline for the introduction of some of the most historically popular compression formats, both for general-purpose data and for specific media types. (Many of these formats actually support multiple distinct compression schemes.) As the timeline illustrates, common compression schemes change every few years, and the explosion of lossy multimedia encoders in the past decade has further accelerated this evolution. This constant churn in popular encoding formats, along with the prevalence of other less common, proprietary or specialized schemes, creates substantial challenges to preserving the usability of digital information over the long term [16].

Figure 1: Timeline of Data Compression Formats
\begin{figure}\centerline{\epsfig{file=timeline.eps, scale=0.22}}\end{figure}

Open compression standards, even when available and widely adopted, do not fully solve these challenges. Specification ambiguities and implementation bugs can make content encoded by one application decode incorrectly or not at all in another. Intellectual property issues such as patents may interfere with the widespread availability of decoders even for ``open'' standards, as occurred in the last decade [4] with several file formats based on the LZW algorithm [33]. Standards also evolve over time, which can make it increasingly difficult to find decoders for obsolete formats that still run on the latest operating systems.

Processor architectures, in contrast, have shown remarkable resistance to change ever since the IBM PC first jump-started personal computing. As the architecture timeline in Figure 2 illustrates, the persistently dominant x86 architecture has experienced only a few major architectural changes during its lifetime--32-bit registers and addressing in 1985, vector processing upgrades starting in 1996, and 64-bit registers and addressing in 2003. More importantly, each of these upgrades has religiously preserved backward code compatibility. Of the other architectures introduced during this period, none have come close to displacing the x86 architecture in the mainstream.

Figure 2: Timeline of Processor Architectures
\begin{figure}\centerline{\epsfig{file=archline.eps, scale=0.22}}\end{figure}

From these facts we observe that instruction encodings are historically more durable than data encodings. We will still be able to run x86 code efficiently decades from now, but it is less likely that future operating systems and applications will still include robust, actively-maintained decoders for today's compressed data streams.

1.1 Virtualizing Decoders

Virtual eXecutable Archives, or VXA, is a novel archival storage architecture that preserves data usability by packaging executable x86-based decoders along with compressed content. These decoders run in a specialized virtual machine (VM) that minimizes dependence on evolving host operating systems and processors. VXA decoders run on a well-defined subset of the unprivileged 32-bit x86 instruction set, and have no direct access to host OS services. A decoder only extracts archived data into simpler, and thus hopefully more ``future-proof,'' uncompressed formats: decoders cannot have user interfaces, open arbitrary files, or communicate with other processes.

By building on the ubiquitous native x86 architecture instead of using a specialized abstract machine such as Lorie's archival ``Universal Virtual Computer'' [27], VXA enables easy re-use of existing decoders written in arbitrary languages such as C and assembly language, which can be built with familiar development tools such as GCC. Use of the x86 architecture also makes execution of virtualized decoders extremely efficient on x86-based host machines, which is important to the many popular ``short-term'' uses of archives such as backups, software distribution, and structured document compression. VXA permits decoders access to the x86 vector processing instructions, further enhancing the performance of multimedia codecs.

Besides preserving long-term data usability, the VXA virtual machine also isolates the host system from buggy or malicious decoders. Decoder security vulnerabilities, such as the recent critical JPEG bug [31], cannot compromise the host under VXA. This security benefit is important because data decoders tend to be inherently complex and difficult to validate, they are frequently exposed to data arriving from untrusted sources such as the Web, and they are usually perceived as too low-level and performance-critical to be written in type-safe languages.

1.2 Prototype Implementation

A prototype implementation of the VXA architecture, vxZIP/vxUnZIP, extends the well-known ZIP/UnZIP archive tools with support for virtualized decoders. The vxZIP archiver can attach VXA decoders both to files it compresses and to input files already compressed with recognized lossy or lossless algorithms. The vxUnZIP archive reader runs these VXA decoders to extract compressed files. Besides enhancing the durability of ZIP files themselves, vxZIP thus also enhances the durability of pre-compressed data stored in ZIP files, and can evolve to employ the latest specialized compression schemes without restricting the usability of the resulting archives.

VXA decoders stored in vxZIP archives are themselves compressed using a fixed algorithm (the ``deflate'' method standard for existing ZIP files) to reduce their storage overhead. The vxZIP prototype currently includes six decoders for both general-purpose data and specialized multimedia streams, ranging from 26 to 130KB in compressed size. Though this storage overhead may be significant for small archives, it is usually negligible for larger archives in which many files share the same decoder.

The prototype vxZIP/vxUnZIP tools run on both the 32-bit and 64-bit variants of the x86 architecture, and rely only on unprivileged facilities available on any mature x86 operating system. The performance cost of virtualization, compared with native x86-32 execution, is between 0 and 11% measured across six widely-available general-purpose and multimedia codecs. The cost is somewhat higher, 8-31%, compared with native x86-64 execution, but this difference is due not to virtualization overhead but to the fact that VXA decoders are always 32-bit, and thus cannot take advantage of the new 64-bit instruction set. The virtual machine that vxUnZIP uses to run the archived decoders is also available as a standalone library, which can be re-used to implement virtualization and isolation of extension modules for other applications.

Section 2 of this paper first presents the VXA architecture in detail. Section 3 then describes the prototype vxZIP/vxUnZIP tools, and Section 4 details the virtual machine monitor in which vxUnZIP runs archived decoders. Section 5 evaluates the performance and storage costs of the virtualized decoders. Finally, Section 6 summarizes related work, and Section 7 concludes.


2 System Architecture

This section introduces the Virtual eXecutable Archive (VXA) architecture at a high level. The principles described in this section are generic and should be applicable to data compression, backup, and archival storage systems of all kinds. All implementation details specific to the prototype VXA archiver and virtual machine are left for the next section.

2.1 Trends and Design Principles

Archived data is almost always compressed in some fashion to save space. The one-time cost of compressing the data in the first place is usually well justified by the savings in storage costs (and perhaps network bandwidth) offered by compression over the long term.

A basic property of data compression, however, is that the more you know about the data being compressed, the more effectively you can compress it. General string-oriented compressors such as gzip do not perform well on digitized photographs, audio, or video, because the information redundancy present in digital media does not predominantly take the form of repeated byte strings, but is specific to the type of media. For this reason a wide variety of media-specific compressors have appeared recently. Lossless compressors achieve moderate compression ratios while preserving all original information content, while lossy compressors achieve higher compression ratios by discarding information whose loss is deemed ``unlikely to be missed'' based on semantic knowledge of the data. Specialization of compression algorithms is not limited to digital media: compressors for semistructured data such as XML are also available for example [26]. This trend toward specialized encodings leads to a first important design principle for efficient archival storage:

An archival storage system must permit use of multiple, specialized compression algorithms.

Strong economic demand for ever more sophisticated and effective data compression has led to a rapid evolution in encoding schemes, even within particular domains such as audio or video, often yielding an abundance of mutually-incompatible competing schemes. Even when open standards achieve widespread use, the dominant standards evolve over time: e.g., from Unix compress to gzip to bzip2. This trend leads to VXA's second basic design principle:

An archival storage system must permit its set of compression algorithms to evolve regularly.

The above two trends unfortunately work against the basic purpose of archival storage: to store data so that it remains available and usable later, perhaps decades later. Even if data is always archived using the latest encoding software, that software--and the operating systems it runs on--may be long obsolete a few years later when the archived data is needed. The widespread use of lossy encoding schemes compounds this problem, because periodically decoding and re-encoding archived data using the latest schemes would cause progressive information loss and thus is not generally a viable option. This constraint leads to VXA's third basic design principle:

Archive extraction must be possible without specific knowledge of the data's encoding.

VXA satisfies these constraints by storing executable decoders with all archived data, and by ensuring that these decoders run in a simple, well-defined, portable, and thus hopefully relatively ``future-proof'' virtual environment.

2.2 Creating Archives

Figure 3 illustrates the basic structure of an archive writer in the VXA architecture. The archiver contains a number of encoder/decoder or codec pairs: several specialized codecs designed to handle specific content types such as audio, video, or XML, and at least one general-purpose lossless codec. The archiver's codec set is extensible via plug-ins, allowing the use of specialized codecs for domain-specific content when desired.

Figure 3: Archive Writer Operation
\begin{figure}\centerline{\epsfig{file=writer.eps, scale=0.30}}\end{figure}

The archiver accepts both uncompressed and already-compressed files as inputs, and automatically tries to compress previously uncompressed input files using a scheme appropriate for the file's type if available. The archiver attempts to compress files of unrecognized type using a general-purpose lossless codec such as gzip. By default the archiver uses only lossless encoding schemes for its automatic compression, but it may apply lossy encoding at the specific request of the operator.

The archiver writes into the archive a copy of the decoder portion of each codec it uses to compress data. The archiver of course needs to include only one copy of a given decoder in the archive, amortizing the storage cost of the decoder over all archived files of that type.

The archiver's codecs can also recognize when an input file is already compressed in a supported format. In this case, the archiver just copies the pre-compressed data into the archive, since re-compressing already-compressed data is generally ineffective and particularly undesirable when lossy compression is involved. The archiver still includes a copy of the appropriate decoder in the archive, ensuring the data's continuing usability even after the original codec has become obsolete or unavailable.

Some of the archiver's codecs may be incapable of compression, but may instead merely recognize files already encoded using other, standalone compressors, and attach a suitable decoder to the archived file. We refer to such pseudo-codecs as recognizer-decoders, or redecs.

2.3 Reading Archives

Figure 4 illustrates the basic structure of the VXA archive reader. Unlike the writer, the reader does not require a collection of content-specific codecs, since all the decoders it needs are embedded in the archive itself. Instead, the archive reader implements a virtual machine in which to run those archived decoders. To decode a compressed file in the archive, the archive reader first locates the associated decoder in the archive and loads it into its virtual machine. The archive reader then executes the decoder in the virtual machine, supplying the encoded data to the decoder while accepting decoded data from the decoder, to produce the decompressed output file.

Figure 4: Archive Reader Operation
\begin{figure}\centerline{\epsfig{file=reader.eps, scale=0.30}}\end{figure}

The archive reader by default only decompresses files that weren't already compressed when the archive was written. This way, archived files that were already compressed in popular standard formats such as JPEG or MP3, which tend to be widely and conveniently usable in their compressed form, remain compressed by default after extraction. The reader can, however, be forced to decode all archived files having an associated decoder, as shown in Figure 4, ensuring that encoded data remains decipherable even if ``native'' decoders for the format disappear.

This capability also helps protect against data corruption caused by codec bugs or evolution of standards. If an archived audio file was generated by a buggy MP3 encoder, for example, it may not play properly later under a different MP3 decoder after extraction from the archive in compressed form. As long as the audio file was originally archived with the specific (buggy) MP3 decoder that can decode the file correctly, however, the archive reader can still be instructed to use that archived decoder to recover a usable decompressed audio stream.

The VXA archive reader does not always have to use the archived x86-based decoders whenever it extracts files from an archive. To maximize performance, the reader might by default recognize popular compressed file types and decode them using non-virtualized decoders compiled for the native host architecture. Such a reader would fall back on running a virtualized decoder from the archive when no suitable native decoder is available, when the native decoder does not work properly on a particular archived stream, or when explicitly checking the archive's integrity. Even if the archive reader commonly uses native rather than virtualized decoders, the presence of the VXA decoders in the archive provides a crucial long-term fallback path for decoding, ensuring that the archived information remains decipherable after the codec it was compressed with has become obsolete and difficult to find.

Routinely using native decoders to read archives instead of the archived VXA decoders, of course, creates the important risk that a bug in a VXA decoder might go unnoticed for a long time, making an archive seem work fine in the short term but be impossible to decode later after the native decoder disappears. For this reason, it is crucial that explicit archive integrity tests always run the archived VXA decoder, and in general it is safest if the archive reader always uses the VXA decoder even when native decoders are available. Since users are unlikely to adopt this safer operational model consistently unless VXA decoder efficiency is on par with native execution, the efficiency of decoder virtualization is more important in practice than it may appear in theory.

2.4 The VXA Virtual Machine

The archive reader's virtual machine isolates the decoders it runs from both the host operating system and the processor architecture on which the archive reader itself runs. Decoders running in the VXA virtual machine have access to the computational primitives of the underlying processor but are extremely limited in terms of input/output. The only I/O decoders are allowed is to read an encoded data stream supplied by the archive reader and produce a corresponding decoded output stream. Decoders cannot access any host operating system services, such as to open files, communicate over the network, or interact with the user.

Through this strong isolation, the virtual machine not only ensures that decoders remain generic and portable across many generations of operating systems, but it also protects the host system from buggy or malicious decoders that may be embedded in an archive. Assuming the virtual machine is implemented correctly, the worst harm a decoder can cause is to garble the data it was supposed to produce from a particular encoded file. Since a decoder cannot communicate, obtain information about the host system, or even check the current system time, decoders do not have access to information with which they might deliberately ``sabotage'' their data based on the conditions under which they are run.

When an archive contains many files associated with the same decoder, the archive reader has the option of re-initializing the virtual machine with a pristine copy of the decoder's executable image before processing each new file, or reusing the virtual machine's state to decode multiple files in succession. Reusing virtual machine state may improve performance, especially on archives containing many small files, at the cost of introducing the risk that a buggy or malicious decoder might ``leak'' information from one file to another during archive extraction, such as from a sensitive password or private key file to a multimedia stream that is likely to appear on a web page. The archive reader can minimize this security risk in practice by always re-initializing the virtual machine whenever the security attributes of the files it is processing change, such as Unix owner/group identifiers and permissions.

The VXA virtual machine is based on the standard 32-bit x86 architecture: all archived decoder executables are represented as x86-32 code, regardless of the actual processor architecture of the host system. The choice of the ubiquitous x86-32 architecture ensures that almost any existing decoder written in any language can be easily ported to run on the VXA virtual machine.

Although continuous improvements in processor hardware are likely to make the performance of an archived VXA decoder largely irrelevant over the long term, compressed archives are frequently used for more short-term purposes as well, such as making and restoring backups, distributing and installing software, and packaging XML-based structured documents [43]. Archive extraction performance is crucial to these short-term uses, and an archival storage system that performs poorly now is unlikely to receive widespread adoption regardless of its long-term benefits. Besides supporting the re-use of existing decoder implementations, VXA's adoption of the x86 architecture also enables those decoders to run quite efficiently on x86-based host processors, as demonstrated later in Section 5. Implementing the VM efficiently on other architectures requires binary translation, which is more difficult and may be less efficient, but is nevertheless by now a practical and proven technology [40,9,14,3].

2.5 Applicability

The VXA architecture does not address the complete problem of preserving the long-term usability of archived digital information. The focus of VXA is on preserving compressed data streams, for which simpler uncompressed formats are readily available that can represent the same information. VXA will not necessarily help with old proprietary word processor documents, for example, for which there is often no obvious ``simpler form'' that preserves all of the original semantic information.

Many document processing applications, however, are moving toward use of ``self-describing'' XML-based structured data formats [43], combined with a general-purpose ``compression wrapper'' such as ZIP [21] for storage efficiency. The VXA architecture may benefit the compression wrapper in such formats, allowing applications to encode documents using proprietary or specialized algorithms for efficiency while preserving the interoperability benefits of XML. VXA's support for specialized compression schemes may be particularly important for XML, in fact, since ``raw'' XML is extremely space-inefficient but can be compressed most effectively given some specialized knowledge of the data [26].


3 Archiver Implementation

Although the basic VXA architecture as described above could be applied to many archival storage or backup systems, the prototype implementation explored in this paper takes the form of an enhancement to the venerable ZIP/UnZIP archival tools [21]. The ZIP format was chosen over the tar/gzip format popular on Unix systems because ZIP compresses files individually rather than as one continuous stream, making it amenable to treating files of different types using different encoders.

For clarity, we will refer to the new VXA-enhanced ZIP and UnZIP utilities here as vxZIP and vxUnZIP, and to the modified archive format as ``vxZIP format.'' In practice, however, the new tools and archive format can be treated as merely a natural upgrade to the existing ones.

3.1 ZIP Archive Format Modifications

The enhanced vxZIP archive format retains the same basic structure and features as the existing ZIP format, and the new utilities remain backward compatible with archives created with existing ZIP tools. Older ZIP tools can list the contents of archives created with vxZIP, but cannot extract files requiring a VXA decoder.

The ZIP file format historically uses a relatively fixed, though gradually growing, collection of general-purpose lossless codecs, each identified by a ``compression method'' tag in a ZIP file. A particular ZIP utility generally compresses all files using only one algorithm by default--the most powerful algorithm it supports--and UnZIP utilities include built-in decoders for most of the compression schemes used by past ZIP utilities. (Decoders for the old LZW-based ``shrinking'' scheme were commonly omitted for many years due to the LZW patent [4], illustrating one of the practical challenges to preserving archived data usability.)

Figure 5: vxZIP Archive Structure
\begin{figure}\centerline{\epsfig{file=format.eps, scale=0.30}}\end{figure}

In the enhanced vxZIP format, an archive may contain files compressed using a mixture of traditional ZIP compression methods and new VXA-specific methods. Files archived using traditional methods are assigned the standard method tag, permitting even VXA-unaware UnZIP tools to identify and extract them successfully. The vxZIP format reserves one new ``special'' ZIP method tag for files compressed using VXA codecs that do not have their own ZIP method tags, and which thus can only be extracted with the help of an attached VXA decoder.

Regardless of whether an archived file uses a traditional or VXA compression scheme, vxZIP attaches a new VXA extension header to each file, pointing to the file's associated VXA decoder, as illustrated in Figure 5. Using this extension header, a VXA-aware archive reader can decode any archived file even if it has an unknown method tag. At the same time, vxUnZIP can still use a file's ZIP method tag to recognize files compressed using well-known algorithms for which it may have a faster native decoder.

When vxZIP recognizes an input file that is already compressed using a scheme for which it has a suitable VXA decoder, it stores the pre-compressed file directly without further compression and tags the file with compression method 0 (no compression). This method tag indicates to vxUnZIP that the file should normally be left compressed on extraction, and enables older UnZIP utilities to extract the file in its original compressed form. The vxZIP archiver nevertheless attaches a VXA decoder to the file in the same way as for automatically-compressed files, so that vxUnZIP can later be instructed to decode the file all the way to its uncompressed form if desired.

3.2 Archiving VXA Decoders

Since the 64KB size limitation of ZIP extension headers precludes storing VXA decoders themselves in the file headers, vxZIP instead stores each decoder elsewhere in the archive as a separate ``pseudo-file'' having its own local file header and an empty filename. The VXA extension headers attached to ``actual'' archived files contain only the ZIP archive offset of the decoder pseudo-file. Many archive files can thus refer to one VXA decoder merely by referring to the same ZIP archive offset.

ZIP archivers write a central directory to the end of each archive, which summarizes the filenames and other meta-data of all files stored in the archive. The vxZIP archiver includes entries in the central directory only for ``actual'' archived files, and not for the pseudo-files containing archived VXA decoders. Since UnZIP tools normally use the central directory when listing the archive's contents, VXA decoder pseudo-files do not show up in such listings even using older VXA-unaware UnZIP tools, and old tools can still use the central directory to find and extract any files not requiring VXA-specific decoders.

A VXA decoder itself is simply an ELF executable for the 32-bit x86 architecture [45], as detailed below in Section 4. VXA decoders are themselves compressed in the archive using a fixed, well-known algorithm: namely the ubiquitous ``deflate'' method used by existing ZIP tools and by the gzip utility popular on Unix systems.

3.3 Codecs for the Archiver

Since a basic goal of the VXA architecture is to be able to support a wide variety of often specialized codecs, it is unacceptable for vxZIP to have a fixed set of built-in compressors, as was generally the case for previous ZIP tools. Instead, vxZIP introduces a plug-in architecture for codecs to be used with the archiver. Each codec consists of two main components:

A natural future extension to this system would be to run VXA encoders as well as decoders in a virtual machine, making complete codec pairs maximally portable. While such an extension should not be difficult, several tradeoffs are involved. A virtual machine for VXA encoders may require user interface support to allow users to configure encoding parameters, introducing additional system complexity. While the performance impact of the VXA virtual machine is not severe at least on x86 hosts, as demonstrated in Section 5, implementing encoders as native DLLs enables the archiving process to run with maximum performance on any host. Finally, vendors of proprietary codecs may not wish to release their encoders for use in a virtualized environment, because it might make license checking more difficult. For these reasons, virtualized VXA encoders are left for future work.


4 The Virtual Machine

The most vital component of the vxUnZIP archive reader is the virtual machine in which it runs archived decoders. This virtual machine is implemented by 32, a novel virtual machine monitor (VMM) that runs in user mode as part of the archive reader's process, without requiring any special privileges or extensions to the host operating system. Decoders under 32 effectively run within vxUnZIP's address space, but in a software-enforced fault isolation domain [46], protecting the application process from possible actions of buggy or malicious decoders. The VMM is implemented as a shared library linked into vxUnZIP; it can also be used to implement specialized virtual machines for other applications.

The 32 VMM currently runs only on x86-based host processors, in both 32-bit and the new 64-bit modes. The VMM relies on quick x86-to-x86 code scanning and translation techniques to sandbox a decoder's code as it executes. These techniques are comparable to those used by Embra [48], VMware [42], and Valgrind [34], though 32 is simpler as it need only provide isolation, and not simulate a whole physical PC or instrument object code for debugging. Full binary translation to make 32 run on other host architectures is under development.

4.1 Data Sandboxing

The VXA virtual machine provides decoders with a ``flat'' unsegmented address space up to 1GB in size, which always starts at virtual address 0 from the perspective of the decoder. The VM does not allow decoders access to the underlying x86 architecture's legacy segmentation facilities. The 32 VMM does, however, use the legacy segmentation features of the x86 host processor in order to implement the virtual machine efficiently.

As illustrated in Figure 6, 32 maps a decoder's virtual address space at some arbitrary location within its own process, and sets up a special process-local (LDT) data segment with a base and limit that provides access only to that region. While running decoder code, the VMM keeps this data segment loaded into the host processor's segment registers that are used for normal data reads and writes (DS, ES, and SS). The decoder's computation and memory access instructions are thus automatically restricted to the sandbox region, without requiring the special code transformations needed on other architectures [46].

Figure 6: Archive Reader and VMM Address Spaces
\begin{figure}\centerline{\epsfig{file=vm.eps, scale=0.30}}\end{figure}

Although the legacy segmentation features that the VMM depends on are not functional in the 64-bit addressing mode (``long mode'') of the new x86-64 processors, these processors provide 64-bit applications the ability to switch back to a 32-bit ``compatibility mode'' in which segmentation features are still available. On a 64-bit system, vxUnZIP and the VMM run in 64-bit long mode, but decoders run in 32-bit compatibility mode. Thus, 32 runs equally well on both x86-32 and x86-64 hosts with only minor implementation differences in the VMM (amounting to about 100 lines of code).


4.2 Code Sandboxing

Although the VMM could similarly set up an x86 code segment that maps only the decoder's address space, doing so would not by itself prevent decoders from executing arbitrary x86 instructions that are ``unsafe'' from the perspective of the VMM, such as those that would modify the segment registers or invoke host operating system calls directly. On RISC-based machines with fixed instruction sizes, a software fault isolation VMM can solve this problem by scanning the untrusted code for ``unsafe'' code sequences when the code is first loaded [46]. This solution is not an option on the x86's variable-length instruction architecture, unfortunately, because within a byte sequence comprising one or more legitimate instructions there may be sub-sequences forming unsafe instructions, to which the decoder code might jump directly. The RISC-based techniques also reserve up to five general-purpose registers as dedicated registers to be used for fault isolation, which is not practical on x86-32 since the architecture provides only eight general-purpose registers total.

The 32 VMM therefore never executes decoder code directly, but instead dynamically scans decoder code sequences to be executed and transforms them into ``safe'' code fragments stored elsewhere in the VMM's process. As with Valgrind [34] and just-in-time compilation techniques [15,24], the VMM keeps transformed code fragments in a cache to be reused whenever the decoder subsequently jumps to the same virtual entrypoint again.

The VMM must of course transform all flow control instructions in the decoder's original code so as to keep execution confined to the safe, transformed code. The VMM rewrites branches with fixed targets to point to the correct transformed code fragment if one already exists. Branches to fixed but as-yet-unknown targets become branches to a ``trampoline'' that, when executed, transforms the target code and then back-patches the original (transformed) branch instruction to point directly to the new target fragment. Finally, the VMM rewrites indirect branches whose target addresses are known only at runtime (including function return instructions), so as to look up the target address dynamically in a hash table of transformed code entrypoints.

4.3 Virtual System Calls

The 32 VMM rewrites x86 instructions that would normally invoke system calls to the host operating system, so as to return control to the user-mode VMM instead. In this way, 32 ensures that decoders have no direct access to host OS services, but can only make controlled ``virtual system calls'' to the VMM or the archive reader.

Only five virtual system calls are available to decoders running under vxUnZIP: read, write, exit, setperm, and done. The first three have their standard Unix meanings, while setperm supports heap memory allocation, and done enables decoders to signal to vxUnZIP that they have finished decoding one stream and are able to process another without being re-loaded. Decoders have access to three standard ``virtual file handles''--stdin, stdout, and stderr--but have no way to open any other files. A decoder's virtual stdin file handle represents the data stream to be decoded, its stdout is the data stream it produces by decoding the input, and stderr serves the traditional purpose of allowing the decoder to write error or debugging messages. (vxUnZIP only displays such messages from decoders when in verbose mode.) A VXA decoder is therefore a traditional Unix filter in a very pure form.

Since a decoder's address space comprises a portion of vxUnZIP's own address space, the archive reader can easily access the decoder's data directly for the purpose of servicing virtual system calls, in the same way that the host OS kernel services system calls made by application processes. To handle the decoder's read and write calls, vxUnZIP merely passes the system call on to the native host OS after checking and adjusting the file handle and buffer pointer arguments. A decoder's I/O calls thus require no extra data copying, and the indirection through the VMM and vxUnZIP code is cheap as it does not cross any hardware protection domains.


5 Evaluation and Results

This section experimentally evaluates the prototype vxZIP/vxUnZIP tools in order to analyze the practicality of the VXA architecture. The two most obvious questions about the practicality of VXA are whether running decoders in a virtual machine seriously compromises their performance for short-term uses of archives such as backups and software/data packaging, and whether embedding decoders in archives entails a significant storage cost. We also consider the portability issues of implementing virtual machines that run x86-32 code on other hosts.

5.1 Test Decoders

The prototype vxZIP archiver includes codecs for several well-known compressed file formats, summarized in Table 1. The two general-purpose codecs, zlib and bzip2, are for arbitrary data streams: vxZIP can use either of them as its ``default compressor'' to compress files of unrecognized type while archiving. The remaining codecs are media-specific. All of the codecs are based directly on publicly-available libraries written in C, and were compiled using a basic GCC cross-compiler setup.


Table 1: Decoders Implemented in vxZIP/vxUnZIP Prototype
Decoder Description Availability Output Format
General-Purpose Codecs
zlib ``Deflate'' algorithm from ZIP/gzip www.zlib.net (raw data)
bzip2 Popular BWT-based algorithm www.bzip.org (raw data)
Still Image Codecs
jpeg Independent JPEG Group (IJG) reference decoder www.ijg.org BMP image
jp2 JPEG-2000 reference decoder from JasPer library www.jpeg.org/jpeg2000 BMP image
Audio Codecs
flac Free Lossless Audio Codec (FLAC) decoder flac.sourceforge.net WAV audio
vorbis Ogg Vorbis audio decoder www.vorbis.com WAV audio


The jpeg and jp2 codecs are recognizer-decoders (``redecs''), which recognize still images compressed in the lossy JPEG [47] and JPEG-2000 [23] formats, respectively, and attach suitable VXA decoders to archived images. These decoders, when run under vxUnZIP, output uncompressed images in the simple and universally-understood Windows BMP file format. The vorbis redec similarly recognizes compressed audio streams in the lossy Ogg/Vorbis format [49], and attaches a Vorbis decoder that yields an uncompressed audio file in the ubiquitous Windows WAV audio file format.

Finally, flac is a full encoder/decoder pair for the Free Lossless Audio Codec (FLAC) format [11]. Using this codec, vxZIP can not only recognize audio streams already compressed in FLAC format and attach a VXA decoder, but it can also recognize uncompressed audio streams in WAV format and automatically compress them using the FLAC encoder. This codec thus demonstrates how a VXA archiver can make use of compression schemes specialized to particular types of data, without requiring the archive reader to contain built-in decoders for each such specialized compression scheme.

The above codecs with widely-available open source implementations were chosen for purposes of evaluating the prototype vxZIP/vxUnZIP implementation, and are not intended to serve as ideal examples to motivate the VXA architecture. While the open formats above may gradually evolve over time, their open-source decoder implementations are unlikely to disappear soon. Commercial archival and multimedia compression products usually incorporate proprietary codecs, however, which might serve as better ``motivating examples'' for VXA: proprietary codecs tend to evolve more quickly due to intense market pressures, and and their closed-source implementations cannot be maintained by the customer or ported to new operating systems once the original product is obsolete and unsupported by the vendor.

5.2 Performance of Virtualized Decoders

To evaluate the performance cost of virtualization, the graph in Figure 7 shows the user-mode CPU time consumed running the above decoders over several test data sets, both natively and under the 32 VMM. All execution times are normalized to the native execution time on an x86-32 host system. The data set used to test the general-purpose lossless codes is a Linux 2.6.11 kernel source tree; the data sets used for the media-specific codecs consist of typical pictures and music files in the appropriate format. All tests were run on an AMD Athlon 64 3000+ with 512MB of RAM, on both the x86-32 and x86-64 versions of SuSE Linux 9.3. The same compiler version (GCC 4.0.0) and optimization settings (-O3) were used for the native and virtualized versions of each decoder, and the timings represent user-mode process time as reported by the time command so as to factor out disk and system overhead. Total wall-clock measurements are not shown because for all but the slowest decoder, jp2, disk overhead dominates total wall-clock time and introduces enough additional variance between successive runs to swamp the differences in CPU-bound decoding time.

Figure 7: Performance of Virtualized Decoders
\begin{figure*}\centerline{\epsfig{file=perf.eps, scale=0.8}}\end{figure*}

As Figure 7 shows, the decoders running under the 32 VMM experience a slowdown of up to 11% relative to native x86-32 execution. The vorbis decoder initially experienced a 29% slowdown when compiled for VXA unmodified, due to subroutine calls in the decoder's inner loop that accentuate the VMM's flow-control overhead by requiring hash table lookups (see Section 4.2). Inlining these two functions both improved the performance of the native decoder slightly (about 1%) and reduced the relative cost of virtualization to 11%. The other decoders were unmodified from their original distribution form. The JPEG decoder became slightly faster under 32, possibly due to effects of the VMM's code rewriting on instruction cache locality; such effects are possible and have been exploited elsewhere [2].

The virtualized decoders fall farther behind in comparison with native execution on an x86-64 host, but this difference is mostly due to the greater efficiency of the 64-bit native code rather than to virtualization overhead. Virtualized decoders always run in 32-bit mode regardless of the host system, so their absolute performance is almost identical on 32-bit versus 64-bit hosts, as the graph shows.

5.3 Decoder Storage Overhead

To evaluate the storage overhead of embedding decoders in archives, Table 2 summarizes the size of each decoder's executable image when compiled and linked for the VXA virtual machine. The code size for each decoder is further split into the portion comprising the decoder itself versus the portion derived from the statically-linked C library against which each decoder is linked. No special effort was made to trim unnecessary code, and the decoders were compiled to optimize performance over code size.


Table 2: Code Size of Virtualized Decoders
Decoder Code Size Compressed
  Total Decoder C Library (zlib)
General-Purpose Codecs  
zlib 46.0KB 32.4KB (70%) 13.6KB (30%) 26.2KB
bzip2 71.1KB 60.9KB (86%) 10.2KB (14%) 29.9KB
Still Image Codecs  
jpeg 103.3KB 90.0KB (87%) 13.3KB (13%) 48.6KB
jp2 220.2KB 198.5KB (90%) 21.7KB (10%) 105.9KB
Audio Codecs  
flac 102.5KB 84.2KB (82%) 18.3KB (18%) 47.6KB
vorbis 233.4KB 200.3KB (86%) 33.1KB (14%) 129.7KB


The significance of these absolute storage overheads of course depends on the size of the archive in which they are embedded, since only one copy of a decoder needs to be stored in the archive regardless of the number of encoded files that use it. As a comparison point, however, a single 2.5-minute CD-quality song in the dataset used for the earlier performance tests, compressed at 120Kbps using the lossy Ogg codec, occupies 2.2MB. The 130KB Ogg decoder for VXA therefore represents a 6% space overhead in an archive containing only this one song, or a 0.6% overhead in an archive containing a 10-song album. The same 2.5-minute song compressed using the lossless FLAC codec occupies 24MB, next to which the 48KB 32 decoder represents a negligible 0.2% overhead.

5.4 Portability Considerations

A clear disadvantage of using the native x86 processor architecture as the basis for VXA decoders is that porting the archive reader to non-x86 host architectures requires instruction set emulation or binary translation. While instruction set emulators can be quite portable, they also tend to be many times slower than native execution, making them unappealing for computation-intensive tasks such as data compression. Binary translation provides better performance and has entered widespread commercial use, but is not simple to implement, and even the best binary translators are unlikely to match the performance of natively-compiled code.

The QEMU x86 emulator [6] introduces a binary translation technique that offers a promising compromise between portability and performance. QEMU uses a native C compiler for the host processor architecture to generate short code fragments that emulate individual x86 instructions. QEMU's dynamic translator then scans the x86 code at run-time and pastes together the appropriate native code fragments to form translated code. While this method is unlikely to perform as well as a binary translator designed and optimized for a specific host architecture, it provides a portable method of implementing emulators that offer usable performance levels.

Even without efficient binary translation for x86 code, however, the cost of emulation does not necessarily make the VXA architecture impractical for non-x86 host architectures. An archive reader can still provide fast native decoders for currently popular file formats, running archived decoders under emulation only when no native decoder is available. The resulting archival system is no slower in practice than existing tools based on a fixed set of compressors, but provides the added assurance that archived data will still be decipherable far into the future. It is much better to be able to decode archived data slowly using emulation than not to be able to decode it at all.

5.5 Availability

The vxZIP/vxUnZIP tools, the 32 virtual machine, and the data sets used in the above tests can be obtained from http://pdos.csail.mit.edu/~baford/vxa/.


6 Related Work

The importance and difficulty of preserving digital information over the long term is gaining increasing recognition [16]. This problem can be broken into two components: preserving data and preserving the data's meaning [13]. Important work is ongoing to address the first aspect [17,12,30], and the second, the focus of this paper, is beginning to receive serious attention.

6.1 Archival Storage Strategies

Storing executable decoders with archived data is not new: popular archivers including ZIP often ship with tools to create self-extracting archives, or executables that decompress themselves when run [35,21]. Such self-extracting archives are designed for convenience, however, and are traditionally specific to a particular host operating system, making them as bad as or worse than traditional non-executable archives for data portability and longevity. Self-extracting archives also provide no security against bugs or malicious decoders; E-mail viruses routinely disguise themselves as self-extracting archives supposedly containing useful applications.

Rothenberg suggested a decade ago the idea of archiving the original application and system software used to create data along with the data itself, and using emulators to run archived software after its original hardware platform becomes obsolete [38]. Archiving entire systems and emulating their hardware accurately is difficult, however, because real hardware platforms (including necessary I/O devices) are extremely complex and tend to be only partly standardized and documented [5]. Preserving the functionality of the original system is also not necessarily equivalent to preserving the usefulness of the original data. The ability to view old data in an emulator window via the original application's archaic user interface, for example, is not the same as being able to load or ``cut-and-paste'' the data into new applications or process it using new indexing or analysis tools.

Lorie later proposed to archive data along with specialized decoder programs, which run on a specialized ``Universal Virtual Computer'' (UVC), and extract archived data into a self-describing XML-like format [27]. The UVC's simplicity makes emulation easier, but since it represents a new architecture substantially different from those of real processors, UVC decoders must effectively be written from scratch in assembly language until high-level languages and tools are developed [28]. More importantly, the UVC's specialization to the ``niche'' of long-term archival storage systems virtually guarantees that high-level languages, development tools, and libraries for it will never be widely available or well-supported as they are for general-purpose architectures.

The LOCKSS archival system supports data format converter plug-ins that transparently migrate data in obsolete formats to new formats when a user accesses the data [37]. Over time, however, actively maintaining converter plug-ins for an ever-growing array of obsolete compressed formats may become difficult. Archiving VXA decoders with compressed data now ensures that future LOCKSS-style ``migrate-on-access'' converters will only need to read common historical uncompressed formats, such as BMP images or WAV audio files, and not the far more numerous and rapidly-evolving compressed formats. VXA therefore complements a ``migrate-on-access'' facility by reducing the number and variety of source formats the access-time converters must support.

6.2 Specialized Virtual Environments

Virtual machines and languages have been designed for many specialized purposes, such as printing [1], boot loading [20], Web programming [19,29], packet filters [32] and other OS extensions [41], active networks [44], active disks [36], and grid computing [8]. In this tradition, VXA could be appropriately described as an architecture for ``active archives.''

Similarly, dynamic code scanning and translation is widely used for purposes such as migrating legacy applications across processor architectures [40,9,3], simulating complete hardware platforms [48], run-time code optimization [2], implementing new processors [14], and debugging compiled code [34,39]. In contrast with the common ``retroactive'' uses of virtual machines and dynamic translation to ``rescue old code'' that no longer runs on the latest systems, however, VXA applies these technologies proactively to preserve the long-term usability and portability of archived data, before the code that knows how to decompress it becomes obsolete.

Most virtual machines designed to support safe application extensions rely on type-safe languages such as Java [7]. In this case, the constraints imposed by the language make the virtual machine more easily portable across processor architectures, at the cost of requiring all untrusted code to be written in such a language. While just-in-time compilation [15,24] has matured to a point where type-safe languages perform adequately for most purposes, some software domains in which performance is traditionally perceived as paramount--such as data compression--remain resolutely attached to unsafe languages such as C and assembly language. Advanced digital media codecs also frequently take advantage of the SIMD extensions of modern processors [22], which tend to be unavailable in type-safe languages. The desire to support the many widespread open and proprietary data encoding algorithms whose implementations are only available in unsafe languages, therefore, makes type-safe language technology infeasible for the VXA architecture.

6.3 Isolation Technologies

The prototype 32 VMM demonstrates a simple and practical software fault isolation (SFI) strategy on the x86, which achieves performance comparable to previous techniques designed for on RISC architectures [46], despite the fact that the RISC-based techniques are not easily applicable to the x86 as discussed in Section 4.2. RISC-based SFI, observed to incur a 15-20% overhead for full virtualization, can be trimmed to 4% overhead by sandboxing memory writes but not reads, thereby protecting the host application from active interference by untrusted code but not from snooping. Unfortunately, this weaker security model is probably not adequate for VXA: a functional but malicious decoder for multimedia files likely to be posted on the Web, for example, could scan the archive reader's address space for data left over from restoring sensitive files such as passwords and private keys from a backup archive, and surreptitiously leak that information into the (public) multimedia output stream it produces.

The Janus security system [18] runs untrusted ``helper'' applications in separate processes, using hardware-based protection in conjunction with Solaris's sophisticated process tracing facilities to control the supervised applications' access to host OS system calls. This approach is more portable across processor architectures than 32's, but less portable across operating systems since it relies on features currently unique to Solaris. The Janus approach also does not enhance the portability of the helper applications, since it does not insulate them from those host OS services they are allowed to access.

The L4 microkernel used an x86-specific segmentation trick analogous to 32's data sandboxing technique to implement fast IPC between small address spaces [25]. A Linux kernel extension similarly used segmentation and paging in combination to give user-level applications a sandbox for untrusted extensions [10]. This latter technique can provide each application with only one virtual sandbox at a time, however, and it imposes constraints on the kernel's own use of x86 segments that would make it impossible to grant use of this facility to 64-bit applications on new x86-64 hosts.


7 Conclusion

The VXA architecture for archival data storage offers a new and practical solution to the problem of preserving the usability of digital content. By including executable decoders in archives that run on a simple and OS-independent virtual machine based on the historically enduring x86 architecture, the VXA architecture ensures that archived data can always be decoded into simpler and less rapidly-evolving uncompressed formats, long after the original codec has become obsolete and difficult to find. The prototype vxZIP/vxUnZIP archiver for x86-based hosts is portable across operating systems, and decoders retain good performance when virtualized.

Acknowledgments

The author wishes to thank Frans Kaashoek, Russ Cox, Maxwell Krohn, and the anonymous reviewers for many helpful comments and suggestions.

Bibliography

1
Adobe Systems Inc.
PostScript Language Reference.
Addison Wesley, 3rd edition, March 1999.

2
Vasanth Bala, Evelyn Duesterwald, and Sanjeev Banerjia.
Dynamo: a transparent dynamic optimization system.
ACM SIGPLAN Notices, 35(5):1-12, 2000.

3
Leonid Baraz, Tevi Devor, Orna Etzion, Shalom Goldenberg, Alex Skaletsky, Yun Wang, and Yigal Zemach.
IA-32 Execution Layer: a two-phase dynamic translator designed to support IA-32 applications on Itanium-based systems.
In 36th International Conference on Microarchitecture (MICRO36), San Diego, CA, December 2003.

4
Michael C. Battilana.
The GIF controversy: A software developer's perspective, June 2004.
http://lzw.info/.

5
David Bearman.
Reality and chimeras in the preservation of electronic records.
D-Lib Magazine, 5(4), April 1999.

6
Fabrice Bellard.
QEMU, a fast and portable dynamic translator, April 2005.

7
Brian Case.
Implementing the Java virtual machine.
Microprocessor Report, 10(4):12-17, March 1996.

8
B. Chang, K. Crary, M. DeLap, R. Harper, J. Liszka, T. Murphy VII, and F. Pfenning.
Trustless grid computing in ConCert.
In Workshop on Grid Computing, pages 112-125, Baltimore, MD, November 2002.

9
A. Chernoff, M. Herdeg, R. Hookway, C. Reeve, N. Rubin, T. Tye, S. Bharadwaj Yadavalli, and J. Yates.
FX!32: a profile-directed binary translator.
IEEE Micro, 18(2):56-64, March 1998.

10
Tzi-cker Chiueh, Ganesh Venkitachalam, and Prashant Pradhan.
Integrating segmentation and paging protection for safe, efficient and transparent software extensions.
In Symposium on Operating System Principles, pages 140-153, December 1999.

11
Josh Coalson.
Free lossless audio codec format.
http://flac.sourceforge.net/format.html.

12
Brian Cooper and Hector Garcia-Molina.
Peer-to-peer data trading to preserve information.
Information Systems, 20(2):133-170, 2002.

13
Arturo Crespo and Hector Garcia-Molina.
Archival storage for digital libraries.
In International Conference on Digital Libraries, 1998.

14
J. Dehnert, B. Grant, J. Banning, R. Johnson, T. Kistler, A. Klaiber, and J. Mattson.
The Transmeta code morphing software: Using speculation, recovery, and adaptive retranslation to address real-life challenges.
In International Symposium on Code Generation and Optimization, pages 15-24. IEEE Computer Society, 2003.

15
L. Peter Deutsch and Allan M. Schiffman.
Efficient implementation of the Smalltalk-80 system.
In Principles of Programming Languages, pages 297-302, Salt Lake City, UT, January 1984.

16
John Garrett and Donald Waters.
Preserving digital information: Report of the task force on archiving of digital information, May 1996.

17
Andrew V. Goldberg and Peter N. Yianilos.
Towards an archival intermemory.
In IEEE Advances in Digital Libraries, pages 147-156, Santa Barbara, CA, 1998. IEEE Computer Society.

18
Ian Goldberg, David Wagner, Randi Thomas, and Eric A. Brewer.
A secure environment for untrusted helper applications.
In 6th USENIX Security Symposium, San Jose, CA, 1996.

19
James Gosling and Henry McGilton.
The Java language environment, May 1996.
http://java.sun.com/docs/white/langenv/.

20
IEEE.
Std 1275-1994: Boot firmware, 1994.

21
Info-ZIP.
http://www.info-zip.org/.

22
Intel Corporation.
IA-32 Intel architecture software developer's manual, June 2005.

23
International Standards Organization.
JPEG 2000 image coding system, 2000.
ISO/IEC 15444-1.

24
Andreas Krall.
Efficient JavaVM just-in-time compilation.
In Parallel Architectures and Compilation Techniques, pages 54-61, Paris, France, October 1998.

25
Jochen Liedtke.
Improved address-space switching on Pentium processors by transparently multiplexing user address spaces.
Technical Report No. 933, GMD -- German National Research Center for Information Technology, November 1995.

26
Hartmut Liefke and Dan Suciu.
XMill: an efficient compressor for XML data.
Technical Report MS-CIS-99-26, University of Pennsylvania, 1999.

27
Raymond A. Lorie.
Long-term archiving of digital information.
Technical Report RJ 10185, IBM Almaden Research Center, May 2000.

28
Raymond A. Lorie.
The UVC: a method for preserving digital documents, proof of concept, 2002.
IBM/KB Long-Term Preservation Study Report Series Number 4.

29
S. Lucco, O. Sharp, and R. Wahbe.
Omniware: A Universal Substrate for Web Programming.
World Wide Web Journal, 1(1):359-368, 1995.

30
Petros Maniatis, Mema Roussopoulos, T. J. Giuli, David S. H. Rosenthal, and Mary Baker.
The LOCKSS peer-to-peer digital preservation system.
Transactions on Computing Systems, 23(1):2-50, 2005.

31
Microsoft Corporation.
Buffer overrun in JPEG processing (GDI+) could allow code execution (833987), September 2004.
Microsoft Security Bulletin MS04-028.

32
Jeffrey C. Mogul, Richard F. Rashid, and Michael J. Accetta.
The packet filter: An efficient mechanism for user-level network code.
In Symposium on Operating System Principles, pages 39-51, Austin, TX, November 1987.

33
Mark Nelson.
LZW data compression.
Dr. Dobb's Journal, October 1989.

34
Nicholas Nethercote and Julian Seward.
Valgrind: A program supervision framework.
In Third Workshop on Runtime Verification (RV'03), Boulder, CO, July 2003.

35
PKWARE Inc.
PKZIP.
http://www.pkware.com/.

36
Erik Riedel, Garth Gibson, and Christos Faloutsos.
Active storage for large-scale data mining and multimedia.
In Very Large Databases (VLDB), New York, NY, August 1998.

37
David S. H. Rosenthal, Thomas Lipkis, Thomas S. Robertson, and Seth Morabito.
Transparent format migration of preserved web content.
D-Lib Magazine, 11(1), January 2005.

38
Jeff Rothenberg.
Ensuring the longevity of digital documents.
Scientific American, 272(1):24-29, January 1995.

39
Julian Seward and Nicholas Nethercote.
Using Valgrind to detect undefined value errors with bit-precision.
In USENIX Annual Technical Conference, Anaheim, CA, April 2005.

40
Richard L. Sites, Anton Chernoff, Matthew B. Kirk, Maurice P. Marks, and Scott G. Robinson.
Binary translation.
Communications of the ACM, 36(2):69-81, 1993.

41
Christopher Small and Margo Seltzer.
A comparison of OS extension technologies.
In USENIX Annual Technical Conference, San Diego, CA, January 1996.

42
Jeremy Sugerman, Ganesh Venkitachalam, and Beng-Hong Lim.
Virtualizing I/O devices on VMware Workstation's hosted virtual machine monitor.
In USENIX Annual Technical Conference, Boston, MA, June 2001.

43
Sun Microsystems.
OpenOffice.org XML file format 1.0, December 2002.
http://xml.openoffice.org.

44
David L. Tennenhouse, Jonathan M. Smith, W. David Sincoskie, David J. Wetherall, and Gary J. Minden.
A survey of active network research.
IEEE Communications Magazine, 35(1):80-86, 1997.

45
Tool Interface Standard (TIS) Committee.
Executable and linking format (ELF) specification, May 1995.

46
Robert Wahbe, Steven Lucco, Thomas E. Anderson, and Susan L. Graham.
Efficient software-based fault isolation.
ACM SIGOPS Operating Systems Review, 27(5):203-216, December 1993.

47
G.K. Wallace.
The JPEG still picture compression standard.
Communications of the ACM, 34(4):30-44, April 1991.

48
Emmett Witchel and Mendel Rosenblum.
Embra: Fast and flexible machine simulation.
In Measurement and Modeling of Computer Systems, pages 68-79, 1996.

49
Xiph.org Foundation.
Vorbis I specification.
http://www.xiph.org/vorbis/.


Bryan Ford