ParPar

ParPar is a high performance, multi-threaded PAR2 creation tool and library for node.js. ParPar does not verify or repair files, only creates redundancy. ParPar is a completely new, from ground-up, implementation, which does not use components from existing PAR2 implementations.

ParPar provides three main things:

• Command line tool for creating PAR2 files, like what par2cmdline does

• High level JS API, with a similar interface to the command line tool

• Low level JS API, providing a high degree of control over the creation process

ParPar is currently still very much a work in progress, is not fully implemented and various issues still need to be solved. Please take this into consideration before using this tool for any non-experimental purposes.

Features

• all main packets from the PAR2 specification

• unicode filename/comment support

• asychronous calculations and I/O

• multi-threading via OpenMP

• multiple fast calculation implementations leveraging x86 (SSE2, SSSE3, AVX2, AVX512BW, GFNI) and ARM (NEON) SIMD capabilities, automatically selecting the best routine for the CPU (see benchmark comparisons)

• multi-buffer (SIMD) MD5 implementation and accelerated CRC32 computation

• single read pass on source files if memory constraints allow (no separate hashing pass required)

• chunking support for memory constrained situations or for generating large amounts of recovery data

• minimum chunk size restrictions to avoid heavy I/O seeking when memory is limited

• cross-platform support

• completely different implementation to all the par2cmdline forks, using fresh new ideas and approaches :)

Possible Future Features

As mentioned above, ParPar is still under development. Some features currently not implemented include:

• improve CLI handling, i.e. better interface

• improve library interface, plus documentation

• better handling of input buffering and processing chunks based on CPU cache size

• improve handling of extremely large (e.g. gigabyte sized) slice sizes

• various other tweaks

Unsupported Features

Here’s a list of features currently not in ParPar, and may never be supported:

• Support for external recovery data or packed slices (I don’t think any PAR2 client supports this)

• Verify/repair PAR2

• Some optimisations in weird edge cases, such as using slice sizes significantly larger than all input data

Motivation

I needed a flexible library for creating PAR2 files, in particular, one which isn’t tied to the notion of on-disk files. ParPar library allows generating PAR2 from “immaterial files”, and the output doesn’t have to be written to disk.

Also, all the fast PAR2 implementations seem to be Windows only; ParPar provides a solution for high performance PAR2 creation on Linux (and probably other) systems (without Wine) and it also works on Windows, as well as non-x86 platforms (i.e. ARM).

Installation / Building

Pre-Built Binaries

Pre-packaged Windows builds with Node 4.x may be found on the releases page if I can be bothered to provide them.

Install Via NPM

If NPM is installed (usually comes bundled with node.js), the following command can be used to install ParPar:

npm install -g @animetosho/parpar

You’ll then be able to run ParPar via the parpar command.

If the npm command isn’t available, it can probably be installed via your package manager (apt-get install npm for Debian), or see the node.js website.

You can then later uninstall ParPar via:

npm uninstall -g @animetosho/parpar

Note that installing from NPM essentially compiles from source, so see issues listed in the following section if the install is failing on the build step.

Install From Source

For building you’ll need node.js (0.10.x or later), node-gyp (can be obtained via npm install -g node-gyp command if NPM is available; may be in package managers otherwise) and relevant build tools (i.e. MS Visual C++ for Windows, GCC/Clang family otherwise). After you have the dependencies, the following commands can be used to build:

node-gyp rebuild
npm install

Note, Windows builds are always compiled with SSE2 support. If you can’t have this, delete all instances of "msvs_settings": {"VCCLCompilerTool": {"EnableEnhancedInstructionSet": "2"}}, in binding.gyp before compiling.

Relatively recent compilers (any supported compiler released in the last ~4 years should work) are needed if AVX support is desired. AVX512 support requires even newer compilers (GCC 5+, MSVC 2017 or later etc).

Building without NPM

If you do not have NPM installed, ParPar can be built easily if your system's package manager has the necessary packages.

Debian 8 is such a system (should also be fine on Ubuntu 14.04), and here's how APT can be used:

apt-get install nodejs node-gyp node-async
node-gyp rebuild

Note that you'll also need node-yencode; follow the instructions here on how to build it. After building it, create a folder named node_modules and place the folder yencode in there.

GCC/Clang build issues

Some versions of GCC/Clang don't like the -march=native switch. If you're having build issues with these compilers, try removing all instances of "-march=native", from binding.gyp and recompiling. Note that some CPU specific optimisations may not be enabled if the flag is removed.

Do also remove the above flag if you are looking to make a portable build.

Multi-Threading Support

ParPar’s multi-threading support requires OpenMP. If ParPar is compiled without OpenMP, it will only ever run on 1 thread. OpenMP is usually available in most compilers that you’d likely use.

It appears that the default compiler in MacOSX does not include OpenMP support (at time of writing). If this is the case, you may need to fetch another build of the C++ compiler which has OpenMP support (e.g. clang/gcc from places like homebrew/macports) and override the CXX environment variable when installing.

API

Note: the terms slices and blocks are used interchangeably here

Simple Example

This is a basic example of the high level JS API (note, API not yet finalised so names etc. may change in future):

var par2creator = require('@animetosho/parpar').run(
    ['file1', 'file2'],   // array of input files
    1024*1024,   // 1MB slice size; if you want a slice count instead, give it as a negative number, e.g. -10 means select a slice size which results in 10 input slices
    {   // options; all these are optional
        outputBase: 'my_recovery_set',
        recoverySlices: { // can also be an array of such objects, of which the sum all these are used
            unit: 'slices', // slices/count, ratio, bytes, largest_files, smallest_files, power or ilog
            value: ,
            scale: 1 // multiply the number of blocks by this amount
        },
        
        // the following are the default values for other options
        //outputBase: '', // output filename without extension
        minSliceSize: null, // default(null) => use sliceSize; give negative number to indicate slice count
        maxSliceSize: null,
        sliceSizeMultiple: 4,
        //recoverySlices: 0,
        minRecoverySlices: null, // default = recoverySlices
        maxRecoverySlices: {
            unit: 'slices',
            value: 65537
        },
        recoveryOffset: 0,
        memoryLimit: 256*1048576,
        minChunkSize: 128*1024, // 0 to disable chunking
        noChunkFirstPass: false,
        processBatchSize: null, // default = max(numthreads * 16, ceil(4M/chunkSize))
        processBufferSize: null, // default = processBatchSize
        comments: [], // array of strings
        unicode: null, // null => auto, false => never, true => always generate unicode packets
        outputOverwrite: false,
        outputIndex: true,
        outputSizeScheme: 'equal', // equal, uniform or pow2
        outputFirstFileSlices: null, // null => default, otherwise pass in same format as outputFileMaxSlices
        outputFirstFileSlicesRounding: 'round', // round, floor or ceil
        outputFileMaxSlices: {
            unit: 'slices',
            value: 65536
        },
        outputFileMaxSlicesRounding: 'round', // round, floor or ceil
        criticalRedundancyScheme: 'pow2', // none or pow2
        outputAltNamingScheme: true,
        displayNameFormat: 'common', // basename, keep or common
        seqReadSize: 4*1048576
    },
    function(err) {
        console.log(err || 'Process finished');
    }
);
par2creator.on('info', function(par) {
    console.log('Creating PAR2 archive with ' + par.opts.recoverySlices*par.opts.sliceSize + ' byte(s) of recovery data from ' + par.totalSize + ' input bytes');
});
par2creator.on('processing_file', function(par, file) {
    console.log('Processing input file ' + file.name);
});
par2creator.on('processing_slice', function(par, file, sliceNum) {
    console.log('Processing slice #' + sliceNum + ' of ' + par.inputSlices + ' from ' + file.name);
});
par2creator.on('pass_complete', function(par, passNum, passChunkNum) {
    console.log('Completed read pass ' + passNum + ' of ' + par.passes + ' pass(es)');
});
par2creator.on('files_written', function(par, passNum, passChunkNum) {
    console.log('Written data for read pass ' + passNum);
});

Functions

Buffer AlignedBuffer(int size)

Returns a normal node Buffer of specified size. The only difference between this and using new Buffer is that this function guarantees the Buffer to be aligned to memory boundaries needed for processing. Use of this function is strictly optional, as ParPar will copy passed input into an AlignedBuffer if the supplied buffer is not aligned. So this function is only useful if you wish to avoid unnecessary memory copying.

 

Remaining API documentation to be done

Examples

Examples for the low level JS API can be found in the examples folder.

Development

Running Tests

Currently only some very basic test scripts are included, which can be found in the aptly named test folder.

md5.js tests the internal MD5 implementation against the reference OpenSSL implementation in Node.

par-compare.js tests PAR2 generation by comparing output from ParPar against that of par2cmdline. As such, par2cmdline needs to be installed for tests to be run. Note that tests will cover extreme cases, including those using large amounts of memory, generating large amounts of recovery data and so on. As such, you will likely need a machine with large amounts of RAM available (preferrably at least 8GB) and reasonable amount of free disk space available (20GB or more recommended) to successfully run all tests.
The test will write several files to a temporary location (sourced from TEMP or TMP environment variables, or the current working directory if none set) and will likely take a while to complete.

Building Binary

Compiling ParPar into a single binary can be done via nexe 1.x. The process is basically the same as building Nyuu’s binary, so see those instructions for details.

GF-Complete

ParPar relies on the excellent GF-Complete library for the heavy lifting. A heavily stripped-down and modified version (from the v2 branch) is included with ParPar. Code from GF-Complete can be found in the gf-complete folder.

Modifications (beyond removing unused components) to the library include:

• MSVC compatibility fixes

• Runtime CPU detection and automatic algorithm selection

• Optimisation for 32-bit builds

• Some optimisations for CPUs without SSE support (SPLIT_TABLE(16,8) implementation)

• Added a Cauchy-like XOR based region multiply technique for faster processing on Atom CPUs, as well as SSE2 CPUs without SSSE3 support

• AVX2 and AVX512BW variants of the SSSE3 (“Shuffle” or SPLIT_TABLE(16,4)) implementation

• Added an experimental technique which is a hybrid of “XOR” and “Shuffle” algorithms above, dubbed “Affine”, which relies on the GF2P8AFFINEQB instruction from GFNI on future Intel processors

• Region transform to/from ALTMAP memory arrangement

• Some tweaks to make ParPar integration easier, such as exposing alignment requirements

Alternatives

For a list of command-line PAR2 tools, see here.

For a nodejs module, there’s node-par2 which is not an implementation of PAR2, rather a wrapper around par2cmdline.

For a C++ library implementation, there’s libpar2, which I believe is based off par2cmdline.

There’s also node-gf, which is a node.js binding for GF-Complete. ParPar’s binding is stripped down for PAR2 purposes, so node-gf is more feature complete and faithful to the original library, but lacks modifications mentioned above.

License

This module is Public Domain.

GF-Complete’s license can be found here.

Multi-buffer MD5 implementation is based off implementation from OpenSSL.