rand: Use stdlib crypto/rand.Read on Linux 6.11 + Go 1.24

[jrick]

On Go 1.24 and Linux >=6.11, the standard library crypto/rand.Read method will read cryptographically secure bytes using the getrandom vDSO. This avoids a context switch to kernel for a getrandom(2) system call and is much faster on small reads than both the previous stdlib crypto/rand reader, and our custom implemented userspace PRNG. It also avoids the need to provide (additional) locking to the dcrd's package global userspace PRNG.

goos: linux
goarch: amd64
pkg: github.com/decred/dcrd/crypto/rand
cpu: Intel(R) Core(TM) i7-14700K
                     │    old.txt    │               new.txt               │
                     │    sec/op     │   sec/op     vs base                │
DcrdRead/4b-20          37.29n ±  2%   34.09n ± 3%   -8.58% (p=0.000 n=10)
DcrdRead/8b-20          41.80n ±  2%   40.41n ± 2%   -3.34% (p=0.000 n=10)
DcrdRead/32b-20         69.45n ±  1%   77.14n ± 2%  +11.07% (p=0.000 n=10)
DcrdRead/512b-20        538.9n ±  2%   674.9n ± 2%  +25.24% (p=0.000 n=10)
DcrdRead/1KiB-20        1.038µ ±  2%   1.308µ ± 1%  +26.01% (p=0.000 n=10)
DcrdRead/4KiB-20        4.069µ ±  2%   5.091µ ± 2%  +25.10% (p=0.000 n=10)
StdlibRead/4b-20       254.25n ±  1%   22.71n ± 2%  -91.07% (p=0.000 n=10)
StdlibRead/8b-20       253.80n ±  1%   29.28n ± 2%  -88.46% (p=0.000 n=10)
StdlibRead/32b-20      253.15n ±  2%   67.43n ± 1%  -73.36% (p=0.000 n=10)
StdlibRead/512b-20      990.3n ±  2%   682.4n ± 2%  -31.09% (p=0.000 n=10)
StdlibRead/1KiB-20      1.717µ ±  4%   1.288µ ± 2%  -24.93% (p=0.000 n=10)
StdlibRead/4KiB-20      6.167µ ±  1%   5.069µ ± 2%  -17.81% (p=0.000 n=10)
DcrdReadPRNG/4b-20      28.86n ±  2%   23.34n ± 2%  -19.12% (p=0.000 n=10)
DcrdReadPRNG/8b-20      33.69n ±  1%   30.17n ± 2%  -10.46% (p=0.000 n=10)
DcrdReadPRNG/32b-20     61.36n ±  1%   68.57n ± 1%  +11.75% (p=0.001 n=10)
DcrdReadPRNG/512b-20    531.4n ±  2%   671.1n ± 2%  +26.30% (p=0.000 n=10)
DcrdReadPRNG/1KiB-20    1.033µ ±  2%   1.312µ ± 1%  +26.96% (p=0.000 n=10)
DcrdReadPRNG/4KiB-20    3.976µ ±  2%   5.086µ ± 2%  +27.90% (p=0.000 n=10)
Int32N-20               46.53n ±  2%   40.80n ± 2%  -12.33% (p=0.000 n=10)
Uint32N-20              46.55n ±  2%   41.51n ± 2%  -10.81% (p=0.000 n=10)
Int64N-20               45.35n ±  2%   41.42n ± 2%   -8.65% (p=0.000 n=10)
Uint64N-20              45.37n ±  3%   41.36n ± 2%   -8.82% (p=0.000 n=10)
Duration-20             46.96n ± 10%   43.34n ± 5%   -7.70% (p=0.002 n=10)
ShuffleSlice-20         38.29n ±  2%   31.05n ± 2%  -18.91% (p=0.000 n=10)
geomean                 208.2n         160.1n       -23.10%

[jrick]

Unfortunately it looks like reading 32 bytes took a small performance hit, unlike the OpenBSD optimization.

But new StdlibRead/32b-20 is faster than old DcrdRead/32b-20? this might need more benchmarking (on a not-loaded system, which I don't have).