README
¶
Reed-Solomon
Introduction:
Erasure Codes(based on Reed-Solomon Codes) engine in pure Go.
It's a kind of Systematic Codes, which means the input data is embedded in the encoded output .
High Performance: More than 15GB/s per physics core.
High Reliability:
- At least two companies are using this library in their storage system. (More than dozens PB data)
- Full test of galois field calculation and invertible matrices (You can also find the mathematical proof in this repo).
Based on Klauspost ReedSolomon & Intel ISA-L with some additional changes/optimizations.
It's the backend of XRS (Erasure Codes which can save about 30% I/O in reconstruction process).
Specification
Math
Coding over in GF(2^8).
Primitive Polynomial: x^8 + x^4 + x^3 + x^2 + 1 (0x1d).
Cauchy Matrix is the generator matrix.
- Any submatrix of encoding matrix is invertible (See the proof here).
Galois Field Tool: Generate primitive polynomial and it's log, exponent, multiply and inverse tables etc.
Inverse Matrices Tool: Calculate the number of inverse matrices with specific data & parity number.
XP has written an excellent article (Here, in Chinese) about how Erasure Codes works and the math behind it. It's a good start to read it.
Accelerate
SIMD: Screaming Fast Galois Field Arithmetic Using Intel SIMD Instructions
Reduce memory I/O: Write cache-friendly code. In the process of two matrices multiply, we will have to read data times, and keep the temporary results, then write to memory. If we could put more data into CPU's Cache but not read/write memory again and again, the performance should improve a lot.
Cache inverse matrices: It'll save thousands ns, not much, but it's still meaningful for small data.
...
Here (in Chinese) is an article about how to write a fast Erasure Codes engine. (Written by me years ago, need update, but the main ideas still work)
Performance
Performance depends mainly on:
CPU instruction extension.
Number of data/parity row vectors.
Platform:
AWS c5d.xlarge (Intel(R) Xeon(R) Platinum 8124M CPU @ 3.00GHz)
All test run on a single Core.
Encode:
I/O = (data + parity) * vector_size / cost
Base means no SIMD.
| Data | Parity | Vector size | AVX512 I/O (MB/S) | AVX2 I/O (MB/S) | Base I/O (MB/S) |
|---|---|---|---|---|---|
| 10 | 2 | 4KB | 29683.69 | 21371.43 | 910.45 |
| 10 | 2 | 1MB | 17664.67 | 15505.58 | 917.26 |
| 10 | 2 | 8MB | 10363.05 | 9323.60 | 914.62 |
| 10 | 4 | 4KB | 17708.62 | 12705.35 | 531.82 |
| 10 | 4 | 1MB | 11970.42 | 9804.57 | 536.31 |
| 10 | 4 | 8MB | 7957.9 | 6941.69 | 534.82 |
| 12 | 4 | 4KB | 16902.12 | 12065.14 | 511.95 |
| 12 | 4 | 1MB | 11478.86 | 9392.33 | 514.24 |
| 12 | 4 | 8MB | 7949.81 | 6760.49 | 513.06 |
Reconstruct:
I/O = (data + reconstruct_data_num) * vector_size / cost
| Data | Parity | Vector size | Reconstruct Data Num | AVX512 I/O (MB/S) |
|---|---|---|---|---|
| 10 | 4 | 4KB | 1 | 29830.36 |
| 10 | 4 | 4KB | 2 | 21649.61 |
| 10 | 4 | 4KB | 3 | 17088.41 |
| 10 | 4 | 4KB | 4 | 14567.26 |
Update:
I/O = (2 + parity_num + parity_num) * vector_size / cost
| Data | Parity | Vector size | AVX512 I/O (MB/S) |
|---|---|---|---|
| 10 | 4 | 4KB | 36444.13 |
Replace:
I/O = (parity_num + parity_num + replace_data_num) * vector_size / cost
| Data | Parity | Vector size | Replace Data Num | AVX512 I/O (MB/S) |
|---|---|---|---|---|
| 10 | 4 | 4KB | 1 | 78464.33 |
| 10 | 4 | 4KB | 2 | 50068.71 |
| 10 | 4 | 4KB | 3 | 38808.11 |
| 10 | 4 | 4KB | 4 | 32457.60 |
| 10 | 4 | 4KB | 5 | 28679.46 |
| 10 | 4 | 4KB | 6 | 26151.85 |
PS:
And we must know the benchmark test is quite different with encoding/decoding in practice. Because in benchmark test loops, the CPU Cache may help a lot.
Links & Thanks
Klauspost ReedSolomon: It's the most commonly used Erasure Codes library in Go. Impressive performance, friendly API, and it can support multi platforms(with fast Galois Field Arithmetic). Inspired me a lot.
Intel ISA-L: The ideas of Cauchy matrix and saving memory I/O are from it.
Documentation
¶
Overview ¶
Warn: Don't modify it. The galois field tables data is generated by mathtool/gentbls.go
Package reedsolomon implements Erasure Codes (systematic codes), it's based on: Reed-Solomon Codes over GF(2^8). Primitive Polynomial: x^8+x^4+x^3+x^2+1.
Galois Filed arithmetic using Intel SIMD instructions (AVX512 or AVX2).
Index ¶
Constants ¶
This section is empty.
Variables ¶
var ( ErrMismatchVects = errors.New("too few/many vects given") ErrZeroVectSize = errors.New("vect size is 0") ErrMismatchVectSize = errors.New("vects size mismatched") )
var ( ErrNoNeedReconst = errors.New("no need reconst") ErrTooManyLost = errors.New("too many lost") ErrHasLostConflict = errors.New("dpHas&lost are conflicting") )
var ( ErrMismatchParityNum = errors.New("parity number mismatched") ErrIllegalVectIndex = errors.New("illegal vect index") )
var ( ErrTooManyReplace = errors.New("too many data for replacing") ErrMismatchReplace = errors.New("number of replaceRows and data mismatch") )
var EnableAVX512 = true
EnableAVX512 may slow down CPU Clock (maybe not). TODO need more research: https://lemire.me/blog/2018/04/19/by-how-much-does-avx-512-slow-down-your-cpu-a-first-experiment/
You can modify it before new RS.
var ErrIllegalVects = errors.New("illegal data/parity number: <= 0 or data+parity > 256")
var ErrNotSquare = errors.New("not a square matrix")
var ErrSingularMatrix = errors.New("matrix is singular")
Functions ¶
func SplitNeedReconst ¶
SplitNeedReconst splits data lost & parity lost.
Types ¶
type RS ¶
type RS struct {
DataNum int // DataNum is the number of data row vectors.
ParityNum int // ParityNum is the number of parity row vectors.
GenMatrix matrix // Generator matrix.
// contains filtered or unexported fields
}
RS Reed-Solomon Codes receiver.
func (*RS) Encode ¶
Encode encodes data for generating parity. It multiplies generator matrix by vects[:r.DataNum] to get parity vectors, and write into vects[r.DataNum:].
func (*RS) Reconst ¶
Reconst reconstructs missing vectors, vects: All vectors, len(vects) = dataNum + parityNum. dpHas: Survived data & parity index, need dataNum indexes at least. needReconst: Vectors indexes which need to be reconstructed.
e.g: in 3+2, the whole index: [0,1,2,3,4], if vects[0,4] are lost & they need to be reconstructed (Maybe you only need vects[0], so the needReconst should be [0], but not [0,4]). the "dpHas" will be [1,2,3] ,and you must be sure that vects[1] vects[2] vects[3] have correct data, results will be written into vects[0]&vects[4] directly.
func (*RS) Replace ¶
Replace replaces oldData vectors with 0 or replaces 0 with newData vectors.
In practice, If len(replaceRows) > dataNum-parityNum, it's better to use Encode, because Replace need to read len(replaceRows) + parityNum vectors, if replaceRows are too many, the cost maybe larger than Encode (Encode only need read dataNum). Think about an EC compute node, and dataNum+parityNum data nodes model.
It's used in two situations: 1. We didn't have enough data for filling in a stripe, but still did ec encode, we need replace several zero vectors with new vectors which have data after we get enough data finally. 2. After compact, we may have several useless vectors in a stripe, we need replaces these useless vectors with zero vectors for free space.
Warn: data's index & replaceRows must has the same sort.
Source Files
Directories
