Wirehair - Fast and Portable Erasure Codes in C
C++ C
Switch branches/tags
Nothing to show
Latest commit 255b9a8 Feb 5, 2014 @chris-gc chris-gc Scale



Fast and Portable Erasure Codes in C

Wirehair produces a stream of error correction blocks from a data source using an erasure code. When enough of these blocks are received, the original data can be recovered. As compared to other similar libraries, an unlimited number of error correction blocks can be produced, and much larger block counts are supported.

A simple C API is provided to make it easy to incorporate into existing projects.

Building: Quick Setup

The wirehair-mobile directory contains an easy-to-import set of C code that also builds properly for mobile devices.

Usage: Encoder

On startup, verify that the Wirehair library is linked correctly:

	if (!wirehair_init()) {
		// Wrong wirehair static library linked

Create an encoder object, providing the message to encode, the number of bytes in the message, and the number of bytes in each block. The message will get broken into equal-sized blocks.

	char *message = ...;
	int bytes = 1000000;
	int block_bytes = 1300;

	wirehair_state encoder = 0;

	encoder = wirehair_encode(0, message, bytes, block_bytes);

	// The encoder object can now be used to write blocks

To check how many blocks are in the message:

	int N = wirehair_count(encoder);

	// N = number of blocks in message

Each time a block will be written it should be assigned a unique ID number, starting from 0 and incremented by one each time:

	char block[1300];
	unsigned int ID = ...;

	if (!wirehair_write(encoder, ID, block)) {

When you are done with the encoder, you can either free the encoder object to reclaim the memory, or reuse the encoder object again. To reuse the object, pass it as the first argument to wirehair_encode. To free the object:


Usage: Decoder

On startup, verify that the Wirehair library is linked correctly as in the encoder case.

A decoder object should be created by providing the size of the message and the number of bytes per block:

	int bytes = 1000000..;
	int block_bytes = 1300;

	wirehair_state decoder = 0;

	decoder = wirehair_decode(0, bytes, block_bytes);

	// The decoder object can now be used to decode the message

To check how many blocks are in the message:

	int N = wirehair_count(decoder);

	// N = number of blocks in message

To decode a message one received block at a time:

	unsigned int ID = ...;

	// If reading is done,
	if (wirehair_read(decoder, ID, block)) {

		char *message = new char[bytes];

		// Decode message
		wirehair_reconstruct(decoder, message);

Note that the wirehair_reconstruct function is used to produce the decoded message. This is suitable for file transfer applications. For packet error correction, a more suitable function is provided:

	char block[1300];
	unsigned int ID = ...;

	wirehair_reconstruct_block(decoder, ID, block);

The Shorthair project is designed to implement efficient variable-length packet error correction using Wirehair on the backend.

Similar to the encoder, you may either free or reuse the decoder object in the exact same way as the encoder.


libwirehair.a on Macbook Air (1.7 GHz Core i5-2557M Sandy Bridge, July 2011):

Turbo Boost is turned on for these computations. The block size is 1300 bytes simulating a file transfer protocol over UDP.

wirehair_encode(N = 12) in 150 usec, 104 MB/s
wirehair_decode(N = 12) average overhead = 0.009 blocks, average reconstruct time = 134.786 usec, 115.739 MB/s

wirehair_encode(N = 32) in 198 usec, 210.101 MB/s
wirehair_decode(N = 32) average overhead = 0.027 blocks, average reconstruct time = 190.22 usec, 218.694 MB/s

wirehair_encode(N = 102) in 310 usec, 427.742 MB/s
wirehair_decode(N = 102) average overhead = 0.016 blocks, average reconstruct time = 378.212 usec, 350.597 MB/s

wirehair_encode(N = 134) in 416 usec, 418.75 MB/s
wirehair_decode(N = 134) average overhead = 0.023 blocks, average reconstruct time = 497.412 usec, 350.213 MB/s

wirehair_encode(N = 169) in 525 usec, 418.476 MB/s
wirehair_decode(N = 169) average overhead = 0.022 blocks, average reconstruct time = 712.778 usec, 308.231 MB/s

wirehair_encode(N = 201) in 654 usec, 399.541 MB/s
wirehair_decode(N = 201) average overhead = 0.019 blocks, average reconstruct time = 798.307 usec, 327.318 MB/s

wirehair_encode(N = 294) in 852 usec, 448.592 MB/s
wirehair_decode(N = 294) average overhead = 0.018 blocks, average reconstruct time = 1048.08 usec, 364.668 MB/s

wirehair_encode(N = 359) in 1176 usec, 396.854 MB/s
wirehair_decode(N = 359) average overhead = 0.021 blocks, average reconstruct time = 1210.73 usec, 385.471 MB/s

wirehair_encode(N = 413) in 1128 usec, 475.975 MB/s
wirehair_decode(N = 413) average overhead = 0.016 blocks, average reconstruct time = 1538.1 usec, 349.067 MB/s

Discussion: Performance Comparison with Reed-Solomon Codes

For values of N under 256, the Wirehair codec exhibits a different performance characteristic as compared to CRS codes (like those in Jerasure):

alt text

The rows are values of k (amount of data) and the columns are values of m (number of recovery blocks added).

Darker is better. The main point of this plot is to just show that m doesn't factor much into the performance of the code.

For comparison, CRS codes have the following type of performance graph (with the same color scale):

alt text

In the case of CRS codes, it is k that doesn't matter.

CRS codes are much simpler, use less memory, and are deterministic, so when they can be used efficiently they are the better option. My implementation of CRS codes is called Longhair.

The issue is that CRS codes are very expensive to use in a lot of cases where Wirehair is more efficient, as shown above.


Wirehair is an FEC codec used to improve reliability of data sent over a Binary Erasure Channel (BEC) such as satellite Internet or WiFi. The data to transmit over the BEC is encoded into equal-sized blocks. When enough blocks are received at the other end of the channel, then the original data can be recovered.

How many additional blocks are needed is random, though the number of additional blocks is low and does not vary based on the size of the data. Typical overhead is 0.03 additional blocks.

Wirehair is designed to be competitive in performance with FEC based on Vandermonde matrices such as zfec, while allowing far more than 256 file blocks (it works best at N=1024 blocks). Furthermore, it can generate a stream of output that goes on forever, which has good recovery properties for any part of that stream that is received, rather than a fixed number of output blocks, making it far more flexible than other libraries.

Discussion: Overhead Reductions with GF(2^16)

Wirehair uses a random 6x18 GF(256) matrix to achieve its low 3% recovery failure rate despite doing most of the calculations on a large sparse binary matrix.

It is trivially possible to decrease the overhead to 0.1% or less by switching to a 9x20 GF(2^16) matrix. I've explored this option and found that it is roughly twice as slow for small block counts, but similar in performance for N >= 1000, since the GF(2^16) matrix does not grow larger.

For worst-case file transfer UDP payloads of roughly 1300 bytes, the average extra bytes needing to be transmitted due to the <3% failure rate is under 40 bytes. Paying a steep performance penalty to reduce this overhead to 4 bytes seems like it is not worth it. For moderately sized messages, 40 bytes is dwarfed by other sources of overhead. If your application needs it, the 16-bit version is available in the source code though it has not been fully tuned.


This software was written entirely by myself ( Christopher A. Taylor mrcatid@gmail.com ). If you find it useful and would like to buy me a coffee, consider tipping.