Commit Graph

5 Commits

Author SHA1 Message Date
Harshavardhana 118270d76f Vendor the upstream changes with Avx512 (#7225)
Thanks to @fwessels we have Avx512 support with 4x improvement
2019-02-12 17:32:27 +05:30
Harshavardhana c0721164be 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)
```
2018-01-03 13:47:22 -08:00
Frank Wessels 6e6aeb6a9e Updated version of klauspost/reedsolomon using proper AVX2 instructions as well a providing support for Cauchy matrices. (#5215) 2017-11-24 14:23:20 -08:00
Frank Wessels 93f126364e Updated version of klauspost/reedsolomon with NEON support for ARM (#4865) 2017-08-30 09:49:00 -07:00
Frank Wessels fffe4ac7e6 Prevent unnecessary verification of parity blocks while reading (#4683)
* Prevent unnecessary verification of parity blocks while reading erasure
  coded file.
* Update klauspost/reedsolomon and just only reconstruct data blocks while
  reading (prevent unnecessary parity block reconstruction)
* Remove Verification of (all) reconstructed Data and Parity blocks since
  in our case we are protected by bit rot protection. And even if the
  verification would fail (essentially impossible) there is no way to
  definitively say whether the data is still correct or not, so this call
  make no sense for our use case.
2017-08-11 18:25:46 -07:00