Memory usage

The memory usage of xz varies from a few hundred kilobytes to several gigabytes depending on the compression settings. The settings used when compressing a file determine the memory requirements of the decompressor. Typically the decompressor needs 5 % to 20 % of the amount of memory that the compressor needed when creating the file. For example, decompressing a file created with xz -9 currently requires 65 MiB of memory. Still, it is possible to have .xz files that require several gigabytes of memory to decompress.

Especially users of older systems may find the possibility of very large memory usage annoying. To prevent uncomfortable surprises, xz has a built-in memory usage limiter, which is disabled by default. While some operating systems provide ways to limit the memory usage of processes, relying on it wasn't deemed to be flexible enough (e.g. using ulimit(1) to limit virtual memory tends to cripple mmap(2)).

The memory usage limiter can be enabled with the command line option --memlimit=limit. Often it is more convenient to enable the limiter by default by setting the environment variable XZ_DEFAULTS, e.g. XZ_DEFAULTS=--memlimit=150MiB. It is possible to set the limits separately for compression and decompression by using --memlimit-compress=limit and --memlimit-decompress=limit. Using these two options outside XZ_DEFAULTS is rarely useful because a single run of xz cannot do both compression and decompression and --memlimit=limit (or -M limit) is shorter to type on the command line.

If the specified memory usage limit is exceeded when decompressing, xz will display an error and decompressing the file will fail. If the limit is exceeded when compressing, xz will try to scale the settings down so that the limit is no longer exceeded (except when using --format=raw or --no-adjust). This way the operation won't fail unless the limit is very small. The scaling of the settings is done in steps that don't match the compression level presets, e.g. if the limit is only slightly less than the amount required for xz -9, the settings will be scaled down only a little, not all the way down to xz -8.

Concatenation and padding with .xz files

It is possible to concatenate .xz files as is. xz will decompress such files as if they were a single .xz file.

It is possible to insert padding between the concatenated parts or after the last part. The padding must consist of null bytes and the size of the padding must be a multiple of four bytes. This can be useful e.g. if the .xz file is stored on a medium that measures file sizes in 512-byte blocks.

Concatenation and padding are not allowed with .lzma files or raw streams.

Integer suffixes and special values

In most places where an integer argument is expected, an optional suffix is supported to easily indicate large integers. There must be no space between the integer and the suffix.

KiB

Multiply the integer by 1,024 (2^10). Ki, k, kB, K, and KB are accepted as synonyms for KiB.

MiB

Multiply the integer by 1,048,576 (2^20). Mi, m, M, and MB are accepted as synonyms for MiB.

GiB

Multiply the integer by 1,073,741,824 (2^30). Gi, g, G, and GB are accepted as synonyms for GiB.

The special value max can be used to indicate the maximum integer value supported by the option.

Operation mode

If multiple operation mode options are given, the last one takes effect.

-z, --compress

Compress. This is the default operation mode when no operation mode option is specified and no other operation mode is implied from the command name (for example, unxz implies --decompress).

-d, --decompress, --uncompress

Decompress.

-t, --test

Test the integrity of compressed files. This option is equivalent to --decompress --stdout except that the decompressed data is discarded instead of being written to standard output. No files are created or removed.

-l, --list

Print information about compressed files. No uncompressed output is produced, and no files are created or removed. In list mode, the program cannot read the compressed data from standard input or from other unseekable sources.

The default listing shows basic information about files, one file per line. To get more detailed information, use also the --verbose option. For even more information, use --verbose twice, but note that this may be slow, because getting all the extra information requires many seeks. The width of verbose output exceeds 80 characters, so piping the output to e.g. less -S may be convenient if the terminal isn't wide enough.

The exact output may vary between xz versions and different locales. For machine-readable output, --robot --list should be used.

Operation modifiers

-k, --keep

Don't delete the input files.

-f, --force

This option has several effects:

-c, --stdout, --to-stdout

Write the compressed or decompressed data to standard output instead of a file. This implies --keep.

--no-sparse

Disable creation of sparse files. By default, if decompressing into a regular file, xz tries to make the file sparse if the decompressed data contains long sequences of binary zeros. It also works when writing to standard output as long as standard output is connected to a regular file and certain additional conditions are met to make it safe. Creating sparse files may save disk space and speed up the decompression by reducing the amount of disk I/O.

-S .suf, --suffix=.suf

When compressing, use .suf as the suffix for the target file instead of .xz or .lzma. If not writing to standard output and the source file already has the suffix .suf, a warning is displayed and the file is skipped.

When decompressing, recognize files with the suffix .suf in addition to files with the .xz, .txz, .lzma, or .tlz suffix. If the source file has the suffix .suf, the suffix is removed to get the target filename.

When compressing or decompressing raw streams (--format=raw), the suffix must always be specified unless writing to standard output, because there is no default suffix for raw streams.

--files[=file]

Read the filenames to process from file; if file is omitted, filenames are read from standard input. Filenames must be terminated with the newline character. A dash (-) is taken as a regular filename; it doesn't mean standard input. If filenames are given also as command line arguments, they are processed before the filenames read from file.

--files0[=file]

This is identical to --files[=file] except that each filename must be terminated with the null character.

Basic file format and compression options

-F format, --format=format

Specify the file format to compress or decompress:

-C check, --check=check

Specify the type of the integrity check. The check is calculated from the uncompressed data and stored in the .xz file. This option has an effect only when compressing into the .xz format; the .lzma format doesn't support integrity checks. The integrity check (if any) is verified when the .xz file is decompressed.

Supported check types:

Integrity of the .xz headers is always verified with CRC32. It is not possible to change or disable it.

-0 ... -9

Select a compression preset level. The default is -6. If multiple preset levels are specified, the last one takes effect. If a custom filter chain was already specified, setting a compression preset level clears the custom filter chain.

The differences between the presets are more significant than with gzip(1) and bzip2(1). The selected compression settings determine the memory requirements of the decompressor, thus using a too high preset level might make it painful to decompress the file on an old system with little RAM. Specifically, it's not a good idea to blindly use -9 for everything like it often is with gzip(1) and bzip2(1).

On the same hardware, the decompression speed is approximately a constant number of bytes of compressed data per second. In other words, the better the compression, the faster the decompression will usually be. This also means that the amount of uncompressed output produced per second can vary a lot.

The following table summarises the features of the presets:

Column descriptions:

-e, --extreme

Use a slower variant of the selected compression preset level (-0 ... -9) to hopefully get a little bit better compression ratio, but with bad luck this can also make it worse. Decompressor memory usage is not affected, but compressor memory usage increases a little at preset levels -0 ... -3.

Since there are two presets with dictionary sizes 4 MiB and 8 MiB, the presets -3e and -5e use slightly faster settings (lower CompCPU) than -4e and -6e, respectively. That way no two presets are identical.

For example, there are a total of four presets that use 8 MiB dictionary, whose order from the fastest to the slowest is -5, -6, -5e, and -6e.

--fast

--best

These are somewhat misleading aliases for -0 and -9, respectively. These are provided only for backwards compatibility with LZMA Utils. Avoid using these options.

--memlimit-compress=limit

Set a memory usage limit for compression. If this option is specified multiple times, the last one takes effect.

If the compression settings exceed the limit, xz will adjust the settings downwards so that the limit is no longer exceeded and display a notice that automatic adjustment was done. Such adjustments are not made when compressing with --format=raw or if --no-adjust has been specified. In those cases, an error is displayed and xz will exit with exit status 1.

The limit can be specified in multiple ways:

See also the section Memory usage.

--memlimit-decompress=limit

Set a memory usage limit for decompression. This also affects the --list mode. If the operation is not possible without exceeding the limit, xz will display an error and decompressing the file will fail. See --memlimit-compress=limit for possible ways to specify the limit.

-M limit, --memlimit=limit, --memory=limit

This is equivalent to specifying --memlimit-compress=limit --memlimit-decompress=limit.

--no-adjust

Display an error and exit if the compression settings exceed the memory usage limit. The default is to adjust the settings downwards so that the memory usage limit is not exceeded. Automatic adjusting is always disabled when creating raw streams (--format=raw).

-T threads, --threads=threads

Specify the number of worker threads to use. The actual number of threads can be less than threads if using more threads would exceed the memory usage limit.

Multithreaded compression and decompression are not implemented yet, so this option has no effect for now.

As of writing (2010-09-27), it hasn't been decided if threads will be used by default on multicore systems once support for threading has been implemented. Comments are welcome. The complicating factor is that using many threads will increase the memory usage dramatically. Note that if multithreading will be the default, it will probably be done so that single-threaded and multithreaded modes produce the same output, so compression ratio won't be significantly affected if threading will be enabled by default.

Custom compressor filter chains

A custom filter chain allows specifying the compression settings in detail instead of relying on the settings associated to the preset levels. When a custom filter chain is specified, the compression preset level options ( -0 ... -9 and --extreme) are silently ignored.

A filter chain is comparable to piping on the command line. When compressing, the uncompressed input goes to the first filter, whose output goes to the next filter (if any). The output of the last filter gets written to the compressed file. The maximum number of filters in the chain is four, but typically a filter chain has only one or two filters.

Many filters have limitations on where they can be in the filter chain: some filters can work only as the last filter in the chain, some only as a non-last filter, and some work in any position in the chain. Depending on the filter, this limitation is either inherent to the filter design or exists to prevent security issues.

A custom filter chain is specified by using one or more filter options in the order they are wanted in the filter chain. That is, the order of filter options is significant! When decoding raw streams (--format=raw), the filter chain is specified in the same order as it was specified when compressing.

Filters take filter-specific options as a comma-separated list. Extra commas in options are ignored. Every option has a default value, so you need to specify only those you want to change.

--lzma1[=options]

--lzma2[=options]

Add LZMA1 or LZMA2 filter to the filter chain. These filters can be used only as the last filter in the chain.

LZMA1 is a legacy filter, which is supported almost solely due to the legacy .lzma file format, which supports only LZMA1. LZMA2 is an updated version of LZMA1 to fix some practical issues of LZMA1. The .xz format uses LZMA2 and doesn't support LZMA1 at all. Compression speed and ratios of LZMA1 and LZMA2 are practically the same.

LZMA1 and LZMA2 share the same set of options:

When decoding raw streams (--format=raw), LZMA2 needs only the dictionary size. LZMA1 needs also lc, lp, and pb.

--x86[=options]

--powerpc[=options]

--ia64[=options]

--arm[=options]

--armthumb[=options]

--sparc[=options]

Add a branch/call/jump (BCJ) filter to the filter chain. These filters can be used only as a non-last filter in the filter chain.

A BCJ filter converts relative addresses in the machine code to their absolute counterparts. This doesn't change the size of the data, but it increases redundancy, which can help LZMA2 to produce 0-15 % smaller .xz file. The BCJ filters are always reversible, so using a BCJ filter for wrong type of data doesn't cause any data loss, although it may make the compression ratio slightly worse.

It is fine to apply a BCJ filter on a whole executable; there's no need to apply it only on the executable section. Applying a BCJ filter on an archive that contains both executable and non-executable files may or may not give good results, so it generally isn't good to blindly apply a BCJ filter when compressing binary packages for distribution.

These BCJ filters are very fast and use insignificant amount of memory. If a BCJ filter improves compression ratio of a file, it can improve decompression speed at the same time. This is because, on the same hardware, the decompression speed of LZMA2 is roughly a fixed number of bytes of compressed data per second.

These BCJ filters have known problems related to the compression ratio:

Both of the above problems will be fixed in the future in a new filter. The old BCJ filters will still be useful in embedded systems, because the decoder of the new filter will be bigger and use more memory.

Different instruction sets have have different alignment:

Since the BCJ-filtered data is usually compressed with LZMA2, the compression ratio may be improved slightly if the LZMA2 options are set to match the alignment of the selected BCJ filter. For example, with the IA-64 filter, it's good to set pb=4 with LZMA2 (2^4=16). The x86 filter is an exception; it's usually good to stick to LZMA2's default four-byte alignment when compressing x86 executables.

All BCJ filters support the same options:

--delta[=options]

Add the Delta filter to the filter chain. The Delta filter can be only used as a non-last filter in the filter chain.

Currently only simple byte-wise delta calculation is supported. It can be useful when compressing e.g. uncompressed bitmap images or uncompressed PCM audio. However, special purpose algorithms may give significantly better results than Delta + LZMA2. This is true especially with audio, which compresses faster and better e.g. with flac(1).

Supported options:

Other options

-q, --quiet

Suppress warnings and notices. Specify this twice to suppress errors too. This option has no effect on the exit status. That is, even if a warning was suppressed, the exit status to indicate a warning is still used.

-v, --verbose

Be verbose. If standard error is connected to a terminal, xz will display a progress indicator. Specifying --verbose twice will give even more verbose output.

The progress indicator shows the following information:

When standard error is not a terminal, --verbose will make xz print the filename, compressed size, uncompressed size, compression ratio, and possibly also the speed and elapsed time on a single line to standard error after compressing or decompressing the file. The speed and elapsed time are included only when the operation took at least a few seconds. If the operation didn't finish, e.g. due to user interruption, also the completion percentage is printed if the size of the input file is known.

-Q, --no-warn

Don't set the exit status to 2 even if a condition worth a warning was detected. This option doesn't affect the verbosity level, thus both --quiet and --no-warn have to be used to not display warnings and to not alter the exit status.

--robot

Print messages in a machine-parsable format. This is intended to ease writing frontends that want to use xz instead of liblzma, which may be the case with various scripts. The output with this option enabled is meant to be stable across xz releases. See the section ROBOT MODE for details.

--info-memory

Display, in human-readable format, how much physical memory (RAM) xz thinks the system has and the memory usage limits for compression and decompression, and exit successfully.

-h, --help

Display a help message describing the most commonly used options, and exit successfully.

-H, --long-help

Display a help message describing all features of xz, and exit successfully

-V, --version

Display the version number of xz and liblzma in human readable format. To get machine-parsable output, specify --robot before --version.

Version

xz --robot --version will print the version number of xz and liblzma in the following format:

XZ_VERSION=XYYYZZZS

LIBLZMA_VERSION=XYYYZZZS

X

Major version.

YYY

Minor version. Even numbers are stable. Odd numbers are alpha or beta versions.

ZZZ

Patch level for stable releases or just a counter for development releases.

S

Stability. 0 is alpha, 1 is beta, and 2 is stable. S should be always 2 when YYY is even.

XYYYZZZS are the same on both lines if xz and liblzma are from the same XZ Utils release.

Examples: 4.999.9beta is 49990091 and 5.0.0 is 50000002.

Memory limit information

xz --robot --info-memory prints a single line with three tab-separated columns:

1.

Total amount of physical memory (RAM) in bytes

2.

Memory usage limit for compression in bytes. A special value of zero indicates the default setting, which for single-threaded mode is the same as no limit.

3.

Memory usage limit for decompression in bytes. A special value of zero indicates the default setting, which for single-threaded mode is the same as no limit.

In the future, the output of xz --robot --info-memory may have more columns, but never more than a single line.

List mode

xz --robot --list uses tab-separated output. The first column of every line has a string that indicates the type of the information found on that line:

name

This is always the first line when starting to list a file. The second column on the line is the filename.

file

This line contains overall information about the .xz file. This line is always printed after the name line.

stream

This line type is used only when --verbose was specified. There are as many stream lines as there are streams in the .xz file.

block

This line type is used only when --verbose was specified. There are as many block lines as there are blocks in the .xz file. The block lines are shown after all the stream lines; different line types are not interleaved.

summary

This line type is used only when --verbose was specified twice. This line is printed after all block lines. Like the file line, the summary line contains overall information about the .xz file.

totals

This line is always the very last line of the list output. It shows the total counts and sizes.

The columns of the file lines:

The columns of the stream lines:

The columns of the block lines:

If --verbose was specified twice, additional columns are included on the block lines. These are not displayed with a single --verbose, because getting this information requires many seeks and can thus be slow:

The columns of the totals line:

If --verbose was specified twice, additional columns are included on the totals line:

Future versions may add new line types and new columns can be added to the existing line types, but the existing columns won't be changed.

Compression preset levels

The numbering of the compression level presets is not identical in xz and LZMA Utils. The most important difference is how dictionary sizes are mapped to different presets. Dictionary size is roughly equal to the decompressor memory usage.

The dictionary size differences affect the compressor memory usage too, but there are some other differences between LZMA Utils and XZ Utils, which make the difference even bigger:

The default preset level in LZMA Utils is -7 while in XZ Utils it is -6, so both use an 8 MiB dictionary by default.

Streamed vs. non-streamed .lzma files

The uncompressed size of the file can be stored in the .lzma header. LZMA Utils does that when compressing regular files. The alternative is to mark that uncompressed size is unknown and use end-of-payload marker to indicate where the decompressor should stop. LZMA Utils uses this method when uncompressed size isn't known, which is the case for example in pipes.

xz supports decompressing .lzma files with or without end-of-payload marker, but all .lzma files created by xz will use end-of-payload marker and have uncompressed size marked as unknown in the .lzma header. This may be a problem in some uncommon situations. For example, a .lzma decompressor in an embedded device might work only with files that have known uncompressed size. If you hit this problem, you need to use LZMA Utils or LZMA SDK to create .lzma files with known uncompressed size.

Unsupported .lzma files

The .lzma format allows lc values up to 8, and lp values up to 4. LZMA Utils can decompress files with any lc and lp, but always creates files with lc=3 and lp=0. Creating files with other lc and lp is possible with xz and with LZMA SDK.

The implementation of the LZMA1 filter in liblzma requires that the sum of lc and lp must not exceed 4. Thus, .lzma files, which exceed this limitation, cannot be decompressed with xz.

LZMA Utils creates only .lzma files which have a dictionary size of 2^n (a power of 2) but accepts files with any dictionary size. liblzma accepts only .lzma files which have a dictionary size of 2^n or 2^n + 2^(n-1). This is to decrease false positives when detecting .lzma files.

These limitations shouldn't be a problem in practice, since practically all .lzma files have been compressed with settings that liblzma will accept.

Trailing garbage

When decompressing, LZMA Utils silently ignore everything after the first .lzma stream. In most situations, this is a bug. This also means that LZMA Utils don't support decompressing concatenated .lzma files.

If there is data left after the first .lzma stream, xz considers the file to be corrupt. This may break obscure scripts which have assumed that trailing garbage is ignored.

Compressed output may vary

The exact compressed output produced from the same uncompressed input file may vary between XZ Utils versions even if compression options are identical. This is because the encoder can be improved (faster or better compression) without affecting the file format. The output can vary even between different builds of the same XZ Utils version, if different build options are used.

The above means that implementing --rsyncable to create rsyncable .xz files is not going to happen without freezing a part of the encoder implementation, which can then be used with --rsyncable.

Embedded .xz decompressors

Embedded .xz decompressor implementations like XZ Embedded don't necessarily support files created with integrity check types other than none and crc32. Since the default is --check=crc64, you must use --check=none or --check=crc32 when creating files for embedded systems.

Outside embedded systems, all .xz format decompressors support all the check types, or at least are able to decompress the file without verifying the integrity check if the particular check is not supported.

XZ Embedded supports BCJ filters, but only with the default start offset.

Basics

Compress the file foo into foo.xz using the default compression level (-6), and remove foo if compression is successful:

Decompress bar.xz into bar and don't remove bar.xz even if decompression is successful:

Create baz.tar.xz with the preset -4e (-4 --extreme), which is slower than e.g. the default -6, but needs less memory for compression and decompression (48 MiB and 5 MiB, respectively):

A mix of compressed and uncompressed files can be decompressed to standard output with a single command:

Parallel compression of many files

On GNU and *BSD, find(1) and xargs(1) can be used to parallelize compression of many files:

The -P option to xargs(1) sets the number of parallel xz processes. The best value for the -n option depends on how many files there are to be compressed. If there are only a couple of files, the value should probably be 1; with tens of thousands of files, 100 or even more may be appropriate to reduce the number of xz processes that xargs(1) will eventually create.

The option -T1 for xz is there to force it to single-threaded mode, because xargs(1) is used to control the amount of parallelization.

Robot mode

Calculate how many bytes have been saved in total after compressing multiple files:

A script may want to know that it is using new enough xz. The following sh(1) script checks that the version number of the xz tool is at least 5.0.0. This method is compatible with old beta versions, which didn't support the --robot option:

Set a memory usage limit for decompression using XZ_OPT, but if a limit has already been set, don't increase it:

Custom compressor filter chains

The simplest use for custom filter chains is customizing a LZMA2 preset. This can be useful, because the presets cover only a subset of the potentially useful combinations of compression settings.

The CompCPU columns of the tables from the descriptions of the options -0 ... -9 and --extreme are useful when customizing LZMA2 presets. Here are the relevant parts collected from those two tables:

If you know that a file requires somewhat big dictionary (e.g. 32 MiB) to compress well, but you want to compress it quicker than xz -8 would do, a preset with a low CompCPU value (e.g. 1) can be modified to use a bigger dictionary:

With certain files, the above command may be faster than xz -6 while compressing significantly better. However, it must be emphasized that only some files benefit from a big dictionary while keeping the CompCPU value low. The most obvious situation, where a big dictionary can help a lot, is an archive containing very similar files of at least a few megabytes each. The dictionary size has to be significantly bigger than any individual file to allow LZMA2 to take full advantage of the similarities between consecutive files.

If very high compressor and decompressor memory usage is fine, and the file being compressed is at least several hundred megabytes, it may be useful to use an even bigger dictionary than the 64 MiB that xz -9 would use:

Using -vv (--verbose --verbose) like in the above example can be useful to see the memory requirements of the compressor and decompressor. Remember that using a dictionary bigger than the size of the uncompressed file is waste of memory, so the above command isn't useful for small files.

Sometimes the compression time doesn't matter, but the decompressor memory usage has to be kept low e.g. to make it possible to decompress the file on an embedded system. The following command uses -6e (-6 --extreme) as a base and sets the dictionary to only 64 KiB. The resulting file can be decompressed with XZ Embedded (that's why there is --check=crc32) using about 100 KiB of memory.

If you want to squeeze out as many bytes as possible, adjusting the number of literal context bits (lc) and number of position bits (pb) can sometimes help. Adjusting the number of literal position bits (lp) might help too, but usually lc and pb are more important. E.g. a source code archive contains mostly US-ASCII text, so something like the following might give slightly (like 0.1 %) smaller file than xz -6e (try also without lc=4):

Using another filter together with LZMA2 can improve compression with certain file types. E.g. to compress a x86-32 or x86-64 shared library using the x86 BCJ filter:

Note that the order of the filter options is significant. If --x86 is specified after --lzma2, xz will give an error, because there cannot be any filter after LZMA2, and also because the x86 BCJ filter cannot be used as the last filter in the chain.

The Delta filter together with LZMA2 can give good results with bitmap images. It should usually beat PNG, which has a few more advanced filters than simple delta but uses Deflate for the actual compression.

The image has to be saved in uncompressed format, e.g. as uncompressed TIFF. The distance parameter of the Delta filter is set to match the number of bytes per pixel in the image. E.g. 24-bit RGB bitmap needs dist=3, and it is also good to pass pb=0 to LZMA2 to accommodate the three-byte alignment:

If multiple images have been put into a single archive (e.g. .tar), the Delta filter will work on that too as long as all images have the same number of bytes per pixel.