Subnostr

I can get about 10k signature verifications per core per second on a 5950X, so about 100us per verification

this is using the unoptimized decred secp256k1 impl and fiatjaf’s go-nostr

oh yeah, sorry, i'm quoting hash/sign/verify here, i haven't actually benched it on here this is a test that signs and verifies using each library, on 9999 events, and probably forming the canonical message and hashing it is part of the operation

i'm really not bothered to test that stuff so much as verify that i've improved upon that shitty implementation, and i have, dramatically, 7x improvement, and since that improvement is likely more than just one thing i've improved, it's probably a net win of more than 20x in total

i am totally pissed at the C and C++ and Rust guys laughing at us Go programmers... on the basis of work done by DECRED devs

fuckin ridiculous

Reply to this note

Please Login to reply.

Discussion

ᴛʜᴇ ᴅᴇᴀᴛʜ ᴏꜰ ᴍʟᴇᴋᴜ 1y ago

yeah, iirc my canonical marshal/hash operation costs about 4us on libsecp256k1 and 11us using go-nostr library, it's retardedly slower... hang on what are the figures i did bench this

https://github.com/mleku/nostrbench

yeah, 1 microsecond for the json op and half that for the binary encoder, and the easyjson is 1.7us idk what the hash function is and that's for the standard object format not the canonical array, which is more complex

my encoders are the fastest you can get in Go and yes i promise not to change them at this point

the downside: part of why it's faster is it isn't using strings, the runtime storage for id, pubkey and sig are raw binary, so it's kinda funny that's still faster than a runtime format that keeps them as hex strings to convert to hex... yes, i use the SIMD xxhex library from templexxx and minio/sha256 hash library for the hash functions, though that's nothing to do with the codec exactly

i'm pretty sure with my codec we have no excuse to not wipe the floor with the repeat myself a thousand time fanbois