slimdata - Man Page
Reversible compression of noisy physics data.
slim [options] filename ...
unslim [options] filename ...
slimcat [options] filename ...
slim performs lossless compression on binary data files. It is designed to operate very rapidly and achieve better compression on noisy physics data than general-purpose tools such as gzip and bzip2.
slim works well only on very specific kinds of data. It requires a file that consists of integers represented in their native 16- or 32-bit binary format and arranged in a regular repetitive pattern. Specifically, slim is designed for the kind of enormous files produced copiously by the Atacama Cosmology Telescope. If it works well for other types of files and from other experiments, then so much the better.
slim is not now appropriate for self-describing files containing their own meta-data (e.g. FITS or NetCDF files). Probably the slim library could be adapted to such files, if a front end were written to parse the meta-data to extract details about the structure of the main data.
Although slim can operate on any file in the sense that the "slimfile" will correctly expand back into the original file, good compression performance--indeed, any compression at all--requires that slim knows specific details about the file contents. See the Data description options on how to describe the file contents, and Raw Data Structure for explanation of the relevant terms.
slim "trains" its encoding algorithms by examining only a fraction of the data. For example, the main algorithm (the reduced binary encoder) works by compressing a small contiguous range of values at the expense of all values outside that range. The ends of that range are selected based on only a sample of the data. For more, see Data sampling options.
unslim is a synonym for slim --expand
slimcat is a synonym for slim --stdout --preserve
All options have short and long forms, which are listed together here.
- -p, --preserve
Preserve the original raw file when compressing or the original slim file when expanding. The default behavior is to delete the input file(s) as each output file is successfully compressed or expanded (except when the option --stdout is used). If this option is selected when expanding files, then the new output file is named by replacing the ".slm" extension with ".raw".
- -X, --compress
The file or files given as arguments are raw files to be compressed. The default is to assume raw files, except where the file names end with the .slm or .SLM suffix. This option is the only way to force compression of a file named something.slm.
- -x, --expand
The file or files given as arguments are slim files to be expanded. The default is to assume raw files, except where the file names end with the .slm or .SLM suffix.
- -k, --force
Force compression or expansion to overwrite any existing file of the same name. Without this option, slim will refuse to overwrite any files.
- -V, --version
Print the slim file version and exit.
- -q, --quiet
Don't print any output unless there are errors.
- -?, --help
Print a usage message and exit.
These options are ignored when expanding files.
- -m, --method=method
Use encoding method method (default = 2). The methods are:
* -m2 The reduced binary encoder (default)
* -m5 Runlength coder. Good for data where the values are
strictly identical for long periods.
All other values are reserved. Methods numbered 1, 3, and 4 were once implemented. They all proved wrong for the job and have been removed. Method 1 was a slight variation on the reduced binary system, differing only in how the parameters were computed. For more on methods 2 through 5, see Encoding Algorithms.
- -d, --deltas
Encoders operate on the differences between successive data values (the deltas). Default is to operate on the data itself. Deltas are slightly slower to encode and decode, but they compress more effectively in almost all situations. In principle, raw data could compress to 0.5 fewer bits per data value, but this would be expected only for extremely stable channels that are not typical of physics data.
- -b, --permit-bitrotation
Bit-rotation is an encoding feature useful in some data streams where the lowest n bits are always the same. When using this option, slim will check each channel for the existence of some number of constant lowest bits. If any are found, then the data are compressed by "rotating" the lowest bits to the top of the number and then proceeding with the usual methods. The author finds this useful with some analog-to-digital schemes that have several low-order bits fixed at zero. For more, see Bit rotation below.
- -n, --filename
Save the original filename in the slim file header. By default, slim does not include the filename in the slim file.
- -S, --rawsize
Save the original file's size (in bytes) in the slim file header. (Currently, this is on by default and cannot be turned off.) Note that when expanding, the raw file size is not tested against the size stored in the file header.
- -C, --compute-crc32
On compression, compute a CRC-32 checksum on the raw data in each section and include its value in the slim file. Default is not to compute the CRC. (On expansion, slim will compute and test the checksum if it is present in the slim file, with or without this option set. However, also see --ignore-crc32 )
These options are ignored when compressing files.
- -o, --stdout
Write the results of expansion to stdout instead of to a file or files. This option implies "--preserve" and is equivalent to running slimcat
- -0, --ignore-crc32
On expansion, do not compute and test the CRC-32 checksum, even if one is included in the slim file.
Data description options
The current version of slim requires that all data channels be the same (i.e. of the same raw type, having the same number of repeats per frame, and so forth). Therefore, when given, all of the options in this section apply to all channels. These requirements are not limitations inherent to the slim file specification, but only to the current implementation of slim. In the future, it could be possible to slim a more heterogeneous mix of data channels, if a configuration file parser were to be added. For more on the differences between the current restrictions on compression and the full generality of the expander, see More general slim data structures.
Data types are assumed to be 1-, 2- or 4-byte integers with little-endian byte order. (Single-byte integer types cannot currently be specified at the command line.) There are no immediate plans to add facilities for handling 8-byte integers or floating-point numbers to slim (although command-line arguments have been reserved for the floating-points).
See Raw Data Structure for more information about how the raw data is assumed to be structured, and for the definitions of the terms channels, sections, frames, and repeats.
All data description options apply to compression only; they are all ignored when expanding a slim file.
- -c, --num-chan=nc
The data consist of nc separate data channels.
- -r, --repeats=reps
One frame contains reps consecutive repeated data values for each channel.
- -F, --frames=nf
There are nf frames in each data section. Note that slim is allowed to break up these frames into multiple sections as needed to keep the section size below the hard limit of 16 MB, or for any other reason. However, slim will also enforce a change to a new section at the end of every nf frames with this option.
- -i, --int
- -u, --unsigned
All data are signed or unsigned 32-bit integers (the default is signed 32-bit).
- -s, --short
- -v, --ushort
All data are signed or unsigned 16-bit integers. Some aspects of slim are not implemented for short integers, such as bit rotation. See BUGS.
- -y, --char
All data are signed 8-bit integers.
- -f, --float
- -g, --double
All data are 32-bit or 64-bit floating point numbers using the IEEE-754 standard. Currently no floating-point compression schemes are implemented, and this option is merely reserved for future use. Instead, 32-bit floating point data are treated as signed 32-bit integers, while 64-bit data are stored without compression.
Data sampling options
The main encoder system (the reduced binary method) works by sampling only some fraction of the raw data and assuming that the statistical properties of the sample are adequate to predict the behavior of the entire set. The data sampling options govern what fraction of the data are used in the sampling. These control the trade-off between speed and compression ratio.
Experience suggests that sampling a greater fraction of the data offers very little benefit; we suggest sampling no more than 10% of all values (the default), unless you plan to do (or have done) a careful test of speed and compression ratio.
Note that the encoders are all universal, in the sense that all possible values can be encoded (though of course most values will not be compressed to a smaller size). This means that the occasional unusual data value can be encoded, and sparse sampling is not fatal to the resulting slim file.
Sampling can be done on a maximum of 20,000 data values. All sample options refer to percentages of either this maximum or the full number of available values, whichever is smaller. Regardless of the options given, a minimum of 20 data values or all values will be sampled, whichever is smaller.
- -G, --sample-pct=pct
Sample pct percent of the data (between 2 and 100%) up to a maximum of 200*pct samples when determining the encoding parameters. Default is 10% up to 2000 samples.
- -9, --18-pct, --best
Sample 18% of the data (up to 3600 samples) to determine encoding parameters.
- -8, --16-pct
Sample 16% of the data (up to 3200 samples) to determine encoding parameters.
- -7, --14-pct
Sample 14% of the data (up to 2800 samples) to determine encoding parameters.
- -6, --12-pct
Sample 12% of the data (up to 2400 samples) to determine encoding parameters.
- -5, --10-pct
Sample 10% of the data (up to 2000 samples) to determine encoding parameters. (This is the default.)
- -4, --8-pct
Sample 8% of the data (up to 1600 samples) to determine encoding parameters.
- -3, --6-pct
Sample 6% of the data (up to 1200 samples) to determine encoding parameters.
- -2, --4-pct
Sample 4% of the data (up to 800 samples) to determine encoding parameters.
- -1, --2-pct, --fast
Sample 2% of the data (up to 400 samples) to determine encoding parameters.
Debugging and unimplemented features
- -B, --debug-buffer=bufsize
read or write raw data from a debugging buffer of size bufsize bytes. This is not the optimal way to compress or expand files (though the price is small). This option is only intended to exercise and debug parts of the code that would be used in a planned slimlib library of functions.
Raw Data Structure
The fundamental concept in slimming a raw data file is that of the data channel. One channel would normally correspond to one physical data source, such a single thermometer or encoder or other sensor. Good compression requires that the distribution of the data from a single channel has a single mode, that its statistical properties of be more or less stationary,that the rate of extreme outliers be small (less than a few percent), and so forth.
Data from more than one channel can be mixed together in the raw file, provided that the pattern of changes from one channel to another is repeated consistently. The most general pattern permitted is to have M samples of channel 1, followed by N samples of channel 2, and so forth until all of the channels in the raw data file have appeared. This unit of the raw file, with each channel occurring one time and giving its expect number of samples, is called the frame.
Multiple frames make up a section. The raw data file is assumed to consist of one or more sections concatenated together. The number, type, repetition count, and order of the channels is fixed throughout a single section (and in the current version of the slim executable, there is no way to change these factors between sections, either). A section consists of one or more frames, and it is valid for the last section in a file to end with a fractional frame (this would, we assume, be due to an unexpected event, such as a raw data file being truncated during acquisition or transmission).
When a raw file contains only one channel, there is an ambiguity in whether that channel is being repeated N times within a single frame, or if it is repeated only once in each of N frames. Because the first choice leads to faster execution, slim silently selects it even if the user's command-line options call for the second. Thus the options "--num-chan=1 --repeats=1 --frames=20000" are silently converted to "--num-chan=1 --repeats=20000 --frames=1".
Note that if one or both of the options giving the repeats or the frames are absent, then slim tries to do the smartest possible thing. You probably don't want to let it do that, but you can.
More general slim data structures
The slim executable produces slim files with several restrictions that are not inherent to the definition of the slim file format itself. The expander can expand files meeting the more general specification, but there is no way at this time to construct these more general files.
Here we list some ways that the structure could be less restrictive than the current compression executable permits. For one thing, the data type of different channels need not be the same, and the number of repeats can also be different. The list of channels could change between one section and the next. Also, the slim file specification allows all sections--not only the last--to be of arbitrary size. However, the current implementation of the slim executable does not offer a way to break up the raw data into sections at arbitrary places in the raw file.
Currently, slim enforces a limit that sections be no longer than 16 MB (2^24 bytes), because of the implementation detail that sections must be held in memory when compressing or expanding them. If data comprising what is conceptually one section exceed this limit, then slim silently divides it into two or more sections of nearly equal length and writes them one after another. This fact is not generally relevant to the user, but it does mean that compression will not be very effective if there are channels that repeat only a few times per 16 MB of raw data.
slim uses a few different algorithms for converting raw data into generally smaller data. By "generally", we mean that no system can convert all possible 32-bit values into smaller ones; this is shown by a simple counting argument. The goal of slim is to recognize a small subset of values that appear most often and to map only this subset into codes that require fewer bits. All other data values outside the small subset are expanded into more than their original number of bits. If most or all raw values come from the compressible subset, then the encoded data will be smaller than the raw data.
Note that all encoding methods currently used in slim convert a single value into an exact integer number of bits (unlike range or arithmetic coding). Thus each bit in a slim file is a part of (or all of) the code for a single value in the raw file (unless, of course, it belongs to the file's header data).
The three encoding methods currently implemented are:
- Constant-value encoding
A channel that contains exactly the same value for every instance will be encoded as a constant-value. This system compresses each value to zero bits, apart from recording the constant value in the section header. No command-line option is required for constant-value encoding: all channels will be checked to see if their values are strictly constant, and if so, they will be encoded by this method in preference over all others.
- Run-length encoding
For a channel whose values are strictly constant for long runs but not for an entire data section, run-length encoding is ideal. For example, it works well on a channel storing the integer part of the time (in seconds), if the time is sampled many times per second. This method will store a repeated run of a single value as a (value, count) pair.
Two cautions are: (1) if the data sample used for evaluation shows that run-length encoding will not actually compress the data channel, then that channel will silently switch over to the standard reduced-binary encoding, and (2) the longest possible run is the number of repetitions within a frame--for technical reasons, a "run" cannot cross from one frame to another. If a channel's values do not appear as several successive words in the raw file, then run-length encoding is not a good choice.
- Reduced-binary encoding
This is the default method. The reduced-binary encoder has two parameters: the number of bits N used for storing the "normal" data, and the offset s subtracted from each value before encoding. The idea is to choose the parameters so that most or all values lie in the range [ s, s+(2^N)-2 ] and can therefore be stored using only N bits. The value s+(2^N)-1 is reserved to indicate "Overflow", that the value being encoded did not lie in the normal range. Overflow codes are followed by the raw data value itself, stored in its natural length of 32 or 16 bits. Thus, most 32-bit data are stored as N-bit numbers, while a small fraction require N+32 bits.
The parameters are chosen using the data sample set. All possible values are tested for N, while s is varied to keep the arithmetic mean of the sample data in the middle of the normal range for the given value of N. The choice of N is that which gives the best compression on the sample data set. Note that keeping the mean in the middle of the range might not always be appropriate for every possible distribution, but that's how it is done.
Tested but rejected encoding algorithms
Three other algorithms were implemented into slim and later removed. Each of them improves on the compression ratios of the reduced-binary code, but the improvements are small and come at a price in compression and expansion speed (not to mention program complexity). Development on these algorithms was dropped for this reason, and we mention them in case you are also pondering ideas for improved encoding algorithms in slim.
- Codes A and B
Codes A and B are minor variations on the reduced-binary code. The former changes how the presence of overflows is signaled, and the latter also changes how their overflowing values are stored.
Code A arises from the observation that overflow values might be rare, but they are still more common than any other single value. The idea is to encode most normal values as an N-bit number but the overflow as an m-bit number, where m<N. The result is to make each overflow take up (N-m) fewer bits, while the allowed range of values contains only 2^(N)-2^(N-m) values, slightly fewer than it would have had otherwise. This results in savings of approximately 0.15 to 0.20 bits per channel for normally-distributed data. However, the method was removed from slim because of its CPU cost: decoding each normal value requires first reading m bits, testing whether they are the overflow code, and (if not) then reading in an additional (N-m) bits to find the full N-bit number.
Code B expands on Code A. It starts with the observation that overflow values are not generally taken randomly from the full range of possible values, but instead are most likely found near the allowed range of non-overflowing values. It is therefore beneficial to write the overflowed value not as a full 32-bit number (assuming the raw data are 32-bit numbers), but instead as an N-bit, or (N+1)-bit number. Thus, overflows are encoded as in Code A, followed by a (prefix, value), where the prefix is an exponential Golomb code for the value's size (actually, the size with N subtracted, since the size of an overflowing value is guaranteed to be no less than N). As with Code A, it improves compression by some fraction of one bit per value, but the savings were judged not to be worth the performance penalty.
- Huffman coding
Huffman coding of the raw data clearly accomplished the best compression ratios of all methods. However, it was also the slowest and so was removed from the program.
The approach is to take the data and split it into "upper" and "lower" bits. The upper bits are Huffman-coded, while the lower bits are assumed to be uniformly-distributed random values and are repeated verbatim into the compressed data stream. The split into upper and lower is made such that the data sample contains no more than 127 distinct values for the upper bits. The Huffman tree is built on the sampled frequencies of the values of the upper bits (with an extra symbol to signal overflows). The choice of 127 symbols was motivated by the possibility of storing the code tree with N symbols in only (N+1) bytes, and by the observation that even as few as 32 symbols work just as well on truly Gaussian-distributed data.
The price of decoding Huffman-encoded symbols is steep: we found that decoding took twice as long when the Huffman encoders were used. The problem is that the prefix-free Huffman codes must be read one bit at a time, for there is no way to know their length in advance.
Bit-rotation is an advanced feature that helps in compressing channels where the lowest B bits are constant. This sort of channel violates the usual assumption that data follow a smooth distribution function, and that the lowest few bits are likely to be distributed almost uniformly. The author of slim has certainly found channels in real experiments where (for example) the six least significant bits are always zero. This might happen a case where an IEEE 754 floating-point number with 24 bits in its sign-plus-mantissa was converted to and recorded as an integer.
If such a channel were encoded by the reduced-binary encoder, then the unchanging B lowest bits would be faithfully but wastefully reproduced over and over again. Bit rotation offers a way around this redundancy.
Bit rotation has slim encode such a channel by bit-shifting the raw value by B bits to the right and then copying what were the B lowest bits to the top (most significant bits) of the shifted result. This makes, for example, the value 0xbeef9900 into 0x00beef99 for the case of B=8. The reduced binary encoder can then act on the bit-rotated result, in which the lowest bits of the new value are presumably less repetitive than they were before bit rotation was applied.
The user selects only whether to try bit rotation with the --permit-bitrotation option. The program then decides whether to choose a non-zero value of B based on the sampled data set. The choice will be the number of lowest-order bits that are strictly constant throughout the sample set.
The penalty for trying bit rotation on every channel is modest when compressing--causing roughly a 1.5% increase in instructions executed, in one test. There is of course no penalty in expanding unless bit rotation was actually used.
32- Versus 64-Bit Architectures and Endianness
In order to operate on machines with 32-bit and 64-bit processors, slim uses strictly sized data types where it matters, such as int32_t for signed 32-bit integers. There are no known problems in running it on machines with 32-bit or 64-bit word sizes. The low-level bit operations are performed on words with the same size as the C++ data type long which would normally be the native word size of the build machine. The code works either way, but it is likely most efficient to operate on words of the native size. The author admits to not having verified this hunch empirically. (You can override this choice and force 32 or 64 bit word sizes. You do this when running build_bit_constants during the build process, but you probably don't want to unless cross-compiling.)
Because of the author's prejudice or narrow experience (or the demands of his day job), the assumption that words are in little-endian order is deeply embedded in slim. An ambitious future contributor is welcome to remove this restriction, but it is hard to see how to handle byte-swapping without a performance penalty. For now, slim fails an assertion if run on a big-endian machine.
It would be very reasonable to have slim use environment variables for some things. Feel free to suggest such features; there are none at this time.
First, compress data from a Multi-channel electronics system. The raw files consist of frames 4400 bytes long having 1100 4-byte words. We ignore that a few of the channels are unsigned and treat them all as signed words. Each channel repeats only once in the frame. We want to encode the differences between successive data values and use method 2 (the reduced binary encoder).
slim -c1100 -i -r1 -dm2 raw_mce_file.dat
Next, a file containing only many, many repetitions of a single channel:
slim -c1 -i -dm2 one_channel_only.dat
To uncompress both files:
unslim raw_mce_file.dat.slm one_channel_only.dat.slm
slim_acthk(1), gzip(1), bzip2(1)
Bugs and Unimplemented Features
The -f and -g command-line options for floating-point data are reserved but aren't implemented.
Bit rotation does not work when the data are 8 or 16-bit integers (whether signed or unsigned).
Copyright © 2008 Joseph Fowler, Princeton University.
This work was supported by and done for the benefit of the Atacama Cosmology Telescope collaboration.
Permission is granted to make and distribute verbatim copies of this manual provided the copyright notice and this permission notice are preserved on all copies.
Permission is granted to copy and distribute modified versions of this manual under the conditions for verbatim copying, provided that the entire resulting derived work is distributed under the terms of a permission notice identical to this one.
Permission is granted to copy and distribute translations of this manual into another language, under the above conditions for modified versions, except that this permission notice may be stated in a translation approved by the Free Software Foundation.