Experiments with AES-NI performance in Golang.
These tests support the implementation of AES-CTR mode in assembly (Golang issue #20967). We are interested in performance improvements obtained by encrypting multiple blocks at the same time (with the same key).
Results show massive improvements up to 4 concurrent blocks and inconsistent
or diminishing returns beyond that. Comparisons to functions that purely
contain AESENC calls suggest the other memory and arithmetic operations are
lost in the noise (as you would expect).
$ go test -bench . -benchtime 20s
BenchmarkSingle-4 5000000000 7.08 ns/op 2258.35 MB/s
BenchmarkMulti/2-4 5000000000 7.75 ns/op 4128.89 MB/s
BenchmarkMulti/4-4 2000000000 12.8 ns/op 5007.81 MB/s
BenchmarkMulti/6-4 2000000000 19.2 ns/op 5010.51 MB/s
BenchmarkMulti/8-4 1000000000 25.7 ns/op 4975.29 MB/s
BenchmarkMulti/10-4 1000000000 32.1 ns/op 4987.93 MB/s
BenchmarkMulti/12-4 1000000000 38.4 ns/op 5000.29 MB/s
BenchmarkMulti/14-4 1000000000 44.6 ns/op 5025.15 MB/s
BenchmarkNomem/2-4 5000000000 6.38 ns/op 5018.09 MB/s
BenchmarkNomem/4-4 2000000000 12.8 ns/op 5000.03 MB/s
BenchmarkNomem/6-4 2000000000 19.1 ns/op 5023.91 MB/s
BenchmarkNomem/8-4 1000000000 25.8 ns/op 4963.32 MB/s
BenchmarkNomem/10-4 1000000000 32.0 ns/op 5005.75 MB/s
BenchmarkNomem/12-4 1000000000 38.6 ns/op 4978.14 MB/s
BenchmarkNomem/14-4 1000000000 44.9 ns/op 4989.37 MB/s