• Ilya Tocar's avatar
    crypto/elliptic: reduce allocations on amd64 · 73f284e2
    Ilya Tocar authored
    This is inspired by
    https://blog.cloudflare.com/go-dont-collect-my-garbage/
    This CL adds allocation tracking and parallelizes p256-related benchmarks.
    Amount of allocations can be significantly reduced by marking amd64 asm
    functions as noescape. This exposes a bug in p256MovCond:
    PANDN with memory argument will fault if memory is not aligned, so they
    are replaced with MOVDQU (which is ok with unaligned memory) and
    register version of PANDN.
    
    Results on 88-thread machine (2x 22 cores) below:
    crypto/elliptic:
    name               old time/op    new time/op    delta
    BaseMultP256-88      1.50µs ±11%    1.19µs ± 5%  -20.20%  (p=0.000 n=10+10)
    ScalarMultP256-88    5.47µs ± 5%    3.63µs ±10%  -33.66%  (p=0.000 n=9+10)
    
    name               old alloc/op   new alloc/op   delta
    BaseMultP256-88        800B ± 0%      288B ± 0%  -64.00%  (p=0.000 n=10+10)
    ScalarMultP256-88    2.59kB ± 0%    0.26kB ± 0%  -90.12%  (p=0.000 n=10+10)
    
    name               old allocs/op  new allocs/op  delta
    BaseMultP256-88        13.0 ± 0%       6.0 ± 0%  -53.85%  (p=0.000 n=10+10)
    ScalarMultP256-88      16.0 ± 0%       5.0 ± 0%  -68.75%  (p=0.000 n=10+10)
    
    crypto/ecdsa:
    name              old time/op    new time/op    delta
    SignP256-88         8.63µs ±37%    7.55µs ±38%     ~     (p=0.393 n=10+10)
    VerifyP256-88       13.9µs ± 8%     7.0µs ± 7%  -49.29%  (p=0.000 n=10+9)
    KeyGeneration-88    2.77µs ±11%    2.34µs ±11%  -15.57%  (p=0.000 n=10+10)
    
    name              old alloc/op   new alloc/op   delta
    SignP256-88         4.14kB ± 1%    2.98kB ± 2%  -27.94%  (p=0.000 n=10+10)
    VerifyP256-88       4.47kB ± 0%    0.99kB ± 0%  -77.84%  (p=0.000 n=9+10)
    KeyGeneration-88    1.21kB ± 0%    0.69kB ± 0%  -42.78%  (p=0.000 n=10+10)
    
    name              old allocs/op  new allocs/op  delta
    SignP256-88           47.0 ± 0%      34.0 ± 0%  -27.66%  (p=0.000 n=10+10)
    VerifyP256-88         38.0 ± 0%      17.0 ± 0%  -55.26%  (p=0.000 n=10+10)
    KeyGeneration-88      20.0 ± 0%      13.0 ± 0%  -35.00%  (p=0.000 n=10+10)
    
    On machine with only 4 cores, results are much less impressive:
    around 2% performance gain.
    
    Change-Id: I8a2f8168f83d27ad9ace1b4b1a1e11cb83edf717
    Reviewed-on: https://go-review.googlesource.com/80757
    Run-TryBot: Ilya Tocar <ilya.tocar@intel.com>
    TryBot-Result: Gobot Gobot <gobot@golang.org>
    Reviewed-by: 's avatarIan Lance Taylor <iant@golang.org>
    73f284e2
p256_amd64.go 16.2 KB