mirror of
https://github.com/minio/minio.git
synced 2025-11-10 22:10:12 -05:00
Automatically set goroutines based on shardSize (#5346)
Update reedsolomon library to enable feature to automatically set number of go-routines based on the input shard size, since shard size is sort of a constant in Minio for objects > 10MiB (default blocksize) klauspost reported around 15-20% improvement in performance numbers on older systems such as AVX and SSE3 ``` name old speed new speed delta Encode10x2x10000-8 5.45GB/s ± 1% 6.22GB/s ± 1% +14.20% (p=0.000 n=9+9) Encode100x20x10000-8 1.44GB/s ± 1% 1.64GB/s ± 1% +13.77% (p=0.000 n=10+10) Encode17x3x1M-8 10.0GB/s ± 5% 12.0GB/s ± 1% +19.88% (p=0.000 n=10+10) Encode10x4x16M-8 7.81GB/s ± 5% 8.56GB/s ± 5% +9.58% (p=0.000 n=10+9) Encode5x2x1M-8 15.3GB/s ± 2% 19.6GB/s ± 2% +28.57% (p=0.000 n=9+10) Encode10x2x1M-8 12.2GB/s ± 5% 15.0GB/s ± 5% +22.45% (p=0.000 n=10+10) Encode10x4x1M-8 7.84GB/s ± 1% 9.03GB/s ± 1% +15.19% (p=0.000 n=9+9) Encode50x20x1M-8 1.73GB/s ± 4% 2.09GB/s ± 4% +20.59% (p=0.000 n=10+9) Encode17x3x16M-8 10.6GB/s ± 1% 11.7GB/s ± 4% +10.12% (p=0.000 n=8+10) ```
This commit is contained in:
committed by
kannappanr
parent
b1fb550d5c
commit
c0721164be
34
vendor/github.com/klauspost/reedsolomon/reedsolomon.go
generated
vendored
34
vendor/github.com/klauspost/reedsolomon/reedsolomon.go
generated
vendored
@@ -15,7 +15,10 @@ import (
|
||||
"bytes"
|
||||
"errors"
|
||||
"io"
|
||||
"runtime"
|
||||
"sync"
|
||||
|
||||
"github.com/klauspost/cpuid"
|
||||
)
|
||||
|
||||
// Encoder is an interface to encode Reed-Salomon parity sets for your data.
|
||||
@@ -239,6 +242,33 @@ func New(dataShards, parityShards int, opts ...Option) (Encoder, error) {
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
if r.o.shardSize > 0 {
|
||||
cacheSize := cpuid.CPU.Cache.L2
|
||||
if cacheSize <= 0 {
|
||||
// Set to 128K if undetectable.
|
||||
cacheSize = 128 << 10
|
||||
}
|
||||
p := runtime.NumCPU()
|
||||
|
||||
// 1 input + parity must fit in cache, and we add one more to be safer.
|
||||
shards := 1 + parityShards
|
||||
g := (r.o.shardSize * shards) / (cacheSize - (cacheSize >> 4))
|
||||
|
||||
if cpuid.CPU.ThreadsPerCore > 1 {
|
||||
// If multiple threads per core, make sure they don't contend for cache.
|
||||
g *= cpuid.CPU.ThreadsPerCore
|
||||
}
|
||||
g *= 2
|
||||
if g < p {
|
||||
g = p
|
||||
}
|
||||
|
||||
// Have g be multiple of p
|
||||
g += p - 1
|
||||
g -= g % p
|
||||
|
||||
r.o.maxGoroutines = g
|
||||
}
|
||||
|
||||
// Inverted matrices are cached in a tree keyed by the indices
|
||||
// of the invalid rows of the data to reconstruct.
|
||||
@@ -431,6 +461,8 @@ func (r reedSolomon) codeSomeShardsP(matrixRows, inputs, outputs [][]byte, outpu
|
||||
if do < r.o.minSplitSize {
|
||||
do = r.o.minSplitSize
|
||||
}
|
||||
// Make sizes divisible by 16
|
||||
do = (do + 15) & (^15)
|
||||
start := 0
|
||||
for start < byteCount {
|
||||
if start+do > byteCount {
|
||||
@@ -490,6 +522,8 @@ func (r reedSolomon) checkSomeShardsP(matrixRows, inputs, toCheck [][]byte, outp
|
||||
if do < r.o.minSplitSize {
|
||||
do = r.o.minSplitSize
|
||||
}
|
||||
// Make sizes divisible by 16
|
||||
do = (do + 15) & (^15)
|
||||
start := 0
|
||||
for start < byteCount {
|
||||
if start+do > byteCount {
|
||||
|
||||
Reference in New Issue
Block a user