reed-solomon-simd

Reed-Solomon erasure coding based on Leopard-RS, featuring:

• O(n log n) complexity.
• Entirely written in Rust.
• Runtime selection of best SIMD implementation on both AArch64 (Neon) and x86(-64) (SSSE3 and AVX2) with fallback to plain Rust.
• Any combination of 1 - 32768 original shards with 1 - 32768 recovery shards.
• Up to 65535 original or recovery shards is also possible with following limitations:

Benchmarks

• Single core AVX2 on an AMD Ryzen 5 3600 (Zen 2, 2019).
• On an Apple Silicon M1 CPU throughput is about the same (+-10%).
• MiB/s and GiB/s are w.r.t the total amount of data, i.e. original shards + recovery shards.
  • For decoder this includes missing shards.
• Shards are 1024 bytes.
• Encode benchmark
  • Includes add_original_shard and encode of ReedSolomonEncoder.
• Decode benchmark
  • Has two MiB/s values for 1% and 100% original shard loss, of maximum possible.
  • Provides minimum required amount of shards to decoder.
  • Includes add_original_shard, add_recovery_shard and decode of ReedSolomonDecoder.

I invite you to clone reed-solomon-simd and run your own benchmark:

$ cargo bench main

Simple usage

1. Divide data into equal-sized original shards. Shard size must be even (shard.len() % 2 == 0).
2. Decide how many recovery shards you want.
3. Generate recovery shards with reed_solomon_simd::encode.
4. When some original shards get lost, restore them with reed_solomon_simd::decode.
   • You must provide at least as many shards as there were original shards in total, in any combination of original shards and recovery shards.

Note: This crate does not detect or correct errors within a shard. So if data corruption is a likely scenario, you should include an error detection hash with each shard, and skip feeding the corrupted shards to the decoder. Here are a few suggestions for very fast error detection hashes: CRC32c (4 bytes), HighwayHash (8, 16 or 32 bytes) or xxHash (4, 8 or 16 bytes).

Example

Divide data into 3 original shards of 64 bytes each and generate 5 recovery shards. Assume then that original shards #0 and #2 are lost and restore them by providing 1 original shard and 2 recovery shards.

let original = [
    b"Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do ",
    b"eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut e",
    b"nim ad minim veniam, quis nostrud exercitation ullamco laboris n",
];

let recovery = reed_solomon_simd::encode(
    3, // total number of original shards
    5, // total number of recovery shards
    original, // all original shards
)?;

let restored = reed_solomon_simd::decode(
    3, // total number of original shards
    5, // total number of recovery shards
    [  // provided original shards with indexes
        (1, &original[1]),
    ],
    [  // provided recovery shards with indexes
        (1, &recovery[1]),
        (4, &recovery[4]),
    ],
)?;

assert_eq!(restored[&0], original[0]);
assert_eq!(restored[&2], original[2]);
# Ok::<(), reed_solomon_simd::Error>(())

Basic usage

ReedSolomonEncoder and ReedSolomonDecoder give more control of the encoding/decoding process.

Here's the above example using these instead:

use reed_solomon_simd::{ReedSolomonDecoder, ReedSolomonEncoder};
use std::collections::HashMap;

let original = [
    b"Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do ",
    b"eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut e",
    b"nim ad minim veniam, quis nostrud exercitation ullamco laboris n",
];

let mut encoder = ReedSolomonEncoder::new(
    3, // total number of original shards
    5, // total number of recovery shards
    64, // shard size in bytes
)?;

for shard in original {
    encoder.add_original_shard(shard)?;
}

let result = encoder.encode()?;
let recovery: Vec<_> = result.recovery_iter().collect();

let mut decoder = ReedSolomonDecoder::new(
    3, // total number of original shards
    5, // total number of recovery shards
    64, // shard size in bytes
)?;

decoder.add_original_shard(1, original[1])?;
decoder.add_recovery_shard(1, recovery[1])?;
decoder.add_recovery_shard(4, recovery[4])?;

let result = decoder.decode()?;
let restored: HashMap<_, _> = result.restored_original_iter().collect();

assert_eq!(restored[&0], original[0]);
assert_eq!(restored[&2], original[2]);
# Ok::<(), reed_solomon_simd::Error>(())

Advanced usage

See rate module for advanced encoding/decoding using chosen Engine and Rate.

Benchmarks against other crates

Use cargo run --release --example quick-comparison to run few simple benchmarks against reed-solomon-16, reed-solomon-erasure, reed-solomon-novelpoly and leopard-codec crates.

This crate is the fastest in all cases on my AMD Ryzen 5 3600, except in the case of decoding with about 42 or fewer recovery shards. There's also a one-time initialization (< 10 ms) for computing tables which can dominate at really small data amounts.

Running tests

Some larger tests are marked #[ignore] and are not run with cargo test. Use cargo test -- --ignored to run those.

Safety

The only use of unsafe in this crate is to allow for target specific optimizations in Ssse3, Avx2 and Neon.

Compatibility

Starting from version 3.0.0, shard sizes that are not multiples of 64 are supported. However, if your shard size is a multiple of 64, it remains compatible across all versions.

Credits

This crate is a fork Markus Laire's reed-solomon-16 crate, which in turn is based on Leopard-RS by Christopher A. Taylor.