Benchmarks I ran and what I concluded from them. Every figure names the hardware it came off.
Five concurrent map implementations across eleven core counts on a 48-core EPYC. Reads scale close to linear for all but one, and the ranking reverses once writes are in the mix. Churn throughput drops every time the core count crosses a power of two.
Sixteen hash candidates across every key shape and length I actually use, benchmarked on the box the code deploys to. The length answer was boring. The type answer was not, and neither was the 7x a string-to-bytes cast costs on the way in.
Three things move a Go TLS 1.3 handshake and they are not close in size. Certificate key type is the biggest lever by far, post-quantum key exchange costs a real 25%, and a keygen patch claws back 9 to 15%.
tumult1337