MrCocoaCat / fastchunking

Fast text chunking algorithms for Python

fastchunking Python library

What it is

fastchunking is a Python library that contains efficient and easy-to-use implementations of string chunking algorithms.

It has been developed as part of the work [LS17] at CISPA, Saarland University.

Installation

$ pip install fastchunking

Note

For performance reasons, parts of this library are implemented in C++. Installation from a source distribution, thus, requires availability of a correctly configured C++ compiler.

Usage and Overview

fastchunking provides efficient implementations for different string chunking algorithms, e.g., static chunking (SC) and content-defined chunking (CDC).

Static Chunking (SC)

Static chunking splits a message into fixed-size chunks.

Let us consider a random example message that shall be chunked:

>>> import os
>>> message = os.urandom(1024*1024)

Static chunking is trivial when chunking a single message:

>>> import fastchunking
>>> sc = fastchunking.SC()
>>> chunker = sc.create_chunker(chunk_size=4096)
>>> chunker.next_chunk_boundaries(message)
[4096, 8192, 12288, ...]

A large message can also be chunked in fragments, though:

>>> chunker = sc.create_chunker(chunk_size=4096)
>>> chunker.next_chunk_boundaries(message[:10240])
[4096, 8192]
>>> chunker.next_chunk_boundaries(message[10240:])
[2048, 6144, 10240, ...]

Content-Defined Chunking (CDC)

fastchunking supports content-defined chunking, i.e., chunking of messages into fragments of variable lengths.

Currently, a chunking strategy based on Rabin-Karp rolling hashes is supported.

As a rolling hash computation on plain-Python strings is incredibly slow with any interpreter, most of the computation is performed by a C++ extension which is based on the ngramhashing library by Daniel Lemire, see: https://github.com/lemire/rollinghashcpp

Let us consider a random message that should be chunked:

>>> import os
>>> message = os.urandom(1024*1024)

When using static chunking, we have to specify a rolling hash window size (here: 48 bytes) and an optional seed value that affects the pseudo-random distribution of the generated chunk boundaries.

Despite that, usage is similar to static chunking:

>>> import fastchunking
>>> cdc = fastchunking.RabinKarpCDC(window_size=48, seed=0)
>>> chunker = cdc.create_chunker(chunk_size=4096)
>>> chunker.next_chunk_boundaries(message)
[7475, 10451, 12253, 13880, 15329, 19808, ...]

Chunking in fragments is straightforward:

>>> chunker = cdc.create_chunker(chunk_size=4096)
>>> chunker.next_chunk_boundaries(message[:10240])
[7475]
>>> chunker.next_chunk_boundaries(message[10240:])
[211, 2013, 3640, 5089, 9568, ...]

Multi-Level Chunking (ML-*)

Multiple chunkers of the same type (but with different chunk sizes) can be efficiently used in parallel, e.g., to perform multi-level chunking [LS17].

Again, let us consider a random message that should be chunked:

>>> import os
>>> message = os.urandom(1024*1024)

Usage of multi-level-chunking, e.g., ML-CDC, is easy:

>>> import fastchunking
>>> cdc = fastchunking.RabinKarpCDC(window_size=48, seed=0)
>>> chunk_sizes = [1024, 2048, 4096]
>>> chunker = cdc.create_multilevel_chunker(chunk_sizes)
>>> chunker.next_chunk_boundaries_with_levels(message)
[(1049, 2), (1511, 1), (1893, 2), (2880, 1), (2886, 0), (3701, 0), (4617, 0), (5809, 2), (5843, 0), ...]

The second value in each tuple indicates the highest chunk size that leads to a boundary. Here, the first boundary is a boundary created by the chunker with index 2, i.e., the chunker with 4096 bytes target chunk size.

Only the highest index is output if multiple chunkers yield the same boundary.

Chunk sizes have to be passed in correct order, i.e., from lowest to highest value.

Performance

Computation costs for static chunking are barely measurable: As chunking does not depend on the actual message but only its length, computation costs are essentially limited to a single xrange call.

Content-defined chunking, however, is expensive: The algorithm has to compute hash values for rolling hash window contents at every byte position of the message that is to be chunked. To minimize costs, fastchunking works as follows:

1. The message (fragment) is passed in its entirety to the C++ extension.
2. Chunking is performed within the C++ extension.
3. The resulting list of chunk boundaries is communicated back to Python and converted into a Python list.

Based on a 100 MiB random content, the author measured the following throughput on an Intel Core i7-4770K in a single, non-representative test run using Python 3.5 (Windows x86-64):

chunk size	throughput
64 bytes	118 MiB/s
128 bytes	153 MiB/s
256 bytes	187 MiB/s
512 bytes	206 MiB/s
1024 bytes	221 MiB/s
2048 bytes	226 MiB/s
4096 bytes	231 MiB/s
8192 bytes	234 MiB/s
16384 bytes	233 MiB/s
32768 bytes	234 MiB/s

Testing

fastchunking uses tox for testing, so simply run:

$ tox

References:

[LS17]

(1, 2) Dominik Leibenger and Christoph Sorge (2017). sec-cs: Getting the Most out of Untrusted Cloud Storage. In Proceedings of the 42nd IEEE Conference on Local Computer Networks (LCN 2017), 2017. (Preprint: arXiv:1606.03368)