[chronode]

This is a cool effort but their claims are wildly misleading as their benchmarks aren't fair across the board:

For example:

- BenchmarkVoidPut runs a single O(n) sized transaction

- BenchmarkLMDBPut runs a single O(n) sized transaction

- BenchmarkBadgetPut runs O(n) O(n) sized transactions (!!!)

I made a local change so that all Put benchmarks ran O(n) O(1) sized transactions and the results were quite different: Void was the slowest, followed by LMDB, Bold, LevelDB, then Badger.

I'd also wager the LMDB author would also (lovingly!) tell us we're holding it wrong

[arandomusername]

LMDB is the only one here in C, so the interop is probably what makes LMDB so slow in here

[apatheticonion]

Curious, what overhead is there calling C code from other languages?

[atombender]

Go doesn't use the C calling convention, but has its own growable stack system and goroutine scheduler that maps to goroutines to threads. So a goroutine can't just call a C function directly.

In order to interface with C code safely, Go's runtime has to jump to the system stack and do some additional setup, make the call, and then switch back. (Adding to that, if the call takes too long, this prevents other goroutines on the same OS thread from running, so the scheduler must jump in and move those goroutines to a different thread.)

All of this is expensive, though we are talking about nanoseconds, not milliseconds. Performance is mostly a problem when doing lots of very quick calls (e.g. you're writing a game engine interacting with something like OpenGL) or lots of slow calls (causing scheduler trashing).

[apatheticonion]

Thanks! Is this also the case even Rust consumes a C library?

[atombender]

No, my understanding is that Rust uses normal stacks, and it uses a classic threading model, so aside from async, calling C doesn't need to any runtime stuff.

[steveklabnik]

That is correct, and was a major motivation for dropping green threads way back in 2014.