A Parallel Compression and Deduplication utility
Pcompress is an attempt to revisit Data Compression using unique combinations of existing and some new techniques. Both high compression ratio and performance are key goals along with the ability to leverage all the cores on a multi-core CPU. It also aims to bring to the table scalable, high-throughput Global Deduplication of archival storage. The deduplication capability is also available for single-file compression modes providing very interesting capabilities. Other projects providing some of these features include Lrzip, eXdupe. Full archivers providing some of the similar features include the excellent FreeArc and PeaZIP. Pcompress is not an archiver but provides a unique combination of features to both maximize compression ratio and provide high speed.
Other open-source deduplication software like OpenDedup and LessFS use fixed block dedupe only. Some software like BackupPC does file-level dedupe only (single-instance storage). Of course OpenDedup and LessFS are Fuse based filesystems doing inline dedupe of primary storage while Pcompress is only meant for archival storage as of today.
NOTE: This utility is Not an archiver. It compresses only single files or datastreams. To archive use something else like tar, cpio or pax.
Blog: https://moinakg.wordpress.com/tag/pcompress/.
Releases: http://freecode.com/projects/pcompress
Current Benchmarks
Benchmarks Set #1
Benchmarks Set #2
Older Benchmarks
Benchmarks Part #1
Benchmarks Part #2
http://code.google.com/p/pcompress/downloads/list
To compress a file:
pcompress -c <algorithm> [-l <compress level>] [-s <chunk size>] <file>
Where <algorithm> can be the folowing:
lzfx - Very fast and small algorithm based on LZF.
lz4 - Ultra fast, high-throughput algorithm reaching RAM B/W at level1.
zlib - The base Zlib format compression (not Gzip).
lzma - The LZMA (Lempel-Ziv Markov) algorithm from 7Zip.
lzmaMt - Multithreaded version of LZMA. This is a faster version but
uses more memory for the dictionary. Thread count is balanced
between chunk processing threads and algorithm threads.
bzip2 - Bzip2 Algorithm from libbzip2.
ppmd - The PPMd algorithm excellent for textual data. PPMd requires
at least 64MB X CPUs more memory than the other modes.
libbsc - A Block Sorting Compressor using the Burrows Wheeler Transform
like Bzip2 but runs faster and gives better compression than
Bzip2 (See: libbsc.com).
adapt - Adaptive mode where ppmd or bzip2 will be used per chunk,
depending on heuristics. If at least 50% of the input data is
7-bit text then PPMd will be used otherwise Bzip2.
adapt2 - Adaptive mode which includes ppmd and lzma. If at least 80% of
the input data is 7-bit text then PPMd will be used otherwise
LZMA. It has significantly more memory usage than adapt.
none - No compression. This is only meaningful with -D and -E so Dedupe
can be done for post-processing with an external utility.
<chunk_size> - This can be in bytes or can use the following suffixes:
g - Gigabyte, m - Megabyte, k - Kilobyte.
Larger chunks produce better compression at the cost of memory.
<compress_level> - Can be a number from 0 meaning minimum and 14 meaning
maximum compression.
NOTE: The option "libbsc" uses Ilya Grebnov's block sorting compression library from http://libbsc.com/ . It is only available if pcompress in built with that library. See INSTALL file for details.
To decompress a file compressed using above command:
pcompress -d <compressed file> <target file>
To operate as a pipe, read from stdin and write to stdout:
pcompress -p ...
Attempt Rabin fingerprinting based deduplication on chunks:
pcompress -D ...
pcompress -D -r ... - Do NOT split chunks at a rabin boundary. Default
is to split.
Perform Delta Encoding in addition to Identical Dedup:
pcompress -E ... - This also implies '-D'. This performs Delta Compression
between 2 blocks if they are 40% to 60% similar. The
similarity %age is selected based on the dedupe block
size to balance performance and effectiveness.
pcompress -EE .. - This causes Delta Compression to happen if 2 blocks are
at least 40% similar regardless of block size. This can
effect greater final compression ratio at the cost of
higher processing overhead.
Number of threads can optionally be specified: -t <1 - 256 count>
Other flags:
'-L' - Enable LZP pre-compression. This improves compression ratio of all
algorithms with some extra CPU and very low RAM overhead. Using
delta encoding in conjunction with this may not always be beneficial.
'-P' - Enable Adaptive Delta Encoding. It can improve compresion ratio further
for data containing tables of numerical values especially if those are
in an arithmetic series. In this implementation basic Delta Encoding is
combined with Run-Length encoding and Matrix transpose
NOTE - Both -L and -P can be used together to give maximum benefit on most
datasets.
'-S' <cksum>
- Specify chunk checksum to use: CRC64, BLAKE2-256, BLAKE2-512, SHA256 and
SHA512. Default one is BLAKE2-256. The implementation actually uses SKEIN
512-256. This is 25% slower than simple CRC64 but is many times more
robust than CRC64 in detecting data integrity errors. BLAKE is a
finalist in the NIST SHA-3 standard selection process and is one of
the fastest in the group, especially on x86 platforms. BLAKE2 is faster
than BLAKE and even faster than MD5.
BLAKE2 512-256 is about 60% faster than SHA 512-256 on x64 platforms.
'-F' - Perform Fixed Block Deduplication. This is faster than fingerprinting
based content-aware deduplication in some cases. However this is mostly
usable for disk dumps especially virtual machine images. This generally
gives lower dedupe ratio than content-aware dedupe (-D) and does not
support delta compression.
'-M' - Display memory allocator statistics
'-C' - Display compression statistics
NOTE: It is recommended not to use '-L' with libbsc compression since libbsc uses LZP internally as well.
Encryption flags:
'-e <ALGO>'
Encrypt chunks using the given encryption algorithm. The algo parameter
can be one of AES or SALSA20. Both are used in CTR stream encryption
mode.
The password can be prompted from the user or read from a file. Unique
keys are generated every time pcompress is run even when giving the same
password. Of course enough info is stored in the compresse file so that
the key used for the file can be re-created given the correct password.
Default key length if 256 bits but can be reduced to 128 bits using the
'-k' option.
The Scrypt algorithm from Tarsnap is used
(See: http://www.tarsnap.com/scrypt.html) for generating keys from
passwords. The CTR mode AES mechanism from Tarsnap is also utilized.
'-w <pathname>'
Provide a file which contains the encryption password. This file must
be readable and writable since it is zeroed out after the password is
read.
'-k <key length>'
Specify the key length. Can be 16 for 128 bit keys or 32 for 256 bit
keys. Default value is 32 for 256 bit keys.
NOTE: When using pipe-mode via -p the only way to provide a password is to use '-w'.
Set ALLOCATOR_BYPASS=1 in the environment to avoid using the the built-in allocator. Due to the the way it rounds up an allocation request to the nearest slab the built-in allocator can allocate extra unused memory. In addition you may want to use a different allocator in your environment.
Compress "file.tar" using bzip2 level 6, 64MB chunk size and use 4 threads. In addition perform identity deduplication and delta compression prior to compression.
pcompress -D -E -c bzip2 -l6 -s64m -t4 file.tar
Compress "file.tar" using extreme compression mode of LZMA and a chunk size of of 1GB. Allow pcompress to detect the number of CPU cores and use as many threads.
pcompress -c lzma -l14 -s1g file.tar
Compress "file.tar" using lz4 at max compression with LZ-Prediction pre-processing and encryption enabled. Chunksize is 100M:
pcompress -c lz4 -l3 -e -L -s100m file.tar
LZFX - Ultra Fast, average compression. This algorithm is the fastest overall. Levels: 1 - 5 LZ4 - Very Fast, better compression than LZFX. Levels: 1 - 3 Zlib - Fast, better compression. Levels: 1 - 9 Bzip2 - Slow, much better compression than Zlib. Levels: 1 - 9 Libbsc - A new Block-Sorting compressor similar conceptually to Bzip2 but gives much better compression. Levels: 1 - 9
LZMA - Very slow. Extreme compression. Levels: 1 - 14 Till level 9 it is standard LZMA parameters. Levels 10 - 12 use more memory and higher match iterations so are slower. Levels 13 and 14 use larger dictionaries upto 256MB and really suck up RAM. Use these levels only if you have at the minimum 4GB RAM on your system.
LzmaMt - Extreme compression, faster than plain LZMA as it is multithreaded. Compression ratio is only slightly less than plain LZMA.
PPMD - Slow. Extreme compression for Text, average compression for binary. In addition PPMD decompression time is also high for large chunks. This requires lots of RAM similar to LZMA. Levels: 1 - 14.
Adapt - Very slow synthetic mode. Both Bzip2 and PPMD are tried per chunk and better result selected. Levels: 1 - 14 Adapt2 - Ultra slow synthetic mode. Both LZMA and PPMD are tried per chunk and better result selected. Can give best compression ratio when splitting file into multiple chunks. Levels: 1 - 14 Since both LZMA and PPMD are used together memory requirements are quite extensive especially if you are also using extreme levels above 10. For example with 64MB chunk, Level 14, 2 threads and with or without dedupe, it uses upto 3.5GB physical RAM and requires 6GB of virtual memory space.
It is possible for a single chunk to span the entire file if enough RAM is available. However for adaptive modes to be effective for large files, especially multi-file archives, splitting into chunks is required so that best compression algorithm can be selected for textual and binary portions.
As can be seen from above memory usage can vary greatly based on compression/ pre-processing algorithms and chunk size. A variety of configurations are possible depending on resource availability in the system.
The minimum possible meaningful settings while still giving about 50% compression ratio and very high speed is with the LZFX algorithm with 1MB chunk size and 2 threads:
pcompress -c lzfx -l2 -s1m -t2 <file>
This uses about 6MB of physical RAM (RSS). Earlier versions of the utility before the 0.9 release comsumed much more memory. This was improved in the later versions. When using Linux the virtual memory consumption may appear to be very high but it is just address space usage rather than actual RAM and should be ignored. It is only the RSS that matters. This is a result of the memory arena mechanism in Glibc that improves malloc() performance for multi-threaded applications.