Today in #iocaine news: I'm converting the benchmarks to Criterion, and am liking what I'm seeing. Groups are neat, and the output is easier to make sense of.
Discussion
Loading...
What if I moved the benchmarks out of iocaine, into a separate repo? That would allow me to make the benchmark repo depend on nam-shub-of-enki, and I could run benchmarks against that, too!
In other news, looking at the Criterion graphs:
- looking up matchers via contstants is a massive ~30% win.
- roto is so much faster than Lua, it is not funny: two orders of magnitude on a test that always returns the same verdict, immediately.
Previously, Lua had the advantage of having global variables, so accessing pre-compiled stuff was considerably faster there. But with emulating globals in Roto, that advantage is lost.
Lua retains the advantage of being a more familiar, and less limited language however. And the advantage of Fennel: the number one reason I added a Lua engine in the first place.
WITH THAT SAID, Lua is still fast. It's about on par with garbage generation (which is also very fast). The choice of request handler language will not be the bottleneck.
I'll come back to benchmarks tonight. Gotta ponder about a few things first, before I continue down this path.
"Tonight" turned into "tomorrow", and I'm back to thinking about benches. I will keep some of them in iocaine, afterall, because it makes sense to have local benchmarks for some core functionalities:
- garbage generation (with our without minify)
- QR code generation (png, inline-svg, external-svg)
- regex-based generation
- trivial request handlers for both roto and lua
- particular request handler-specific benches (hashmap/constant lookups for roto)
I can have an external benchmark for more complex stuff.
21 files changed, 414 insertions(+), 732 deletions(-)
So far, so good.
What this is? This is lifting out the various matchers common between the Roto and Lua engines into their own thing, and using thin wrappers from Roto/Lua around that, instead of reimplementing every matcher for both.
It also changes the signature of the Roto functions from function init() -> Verdict[Unit, String] to fn init() -> bool? and function decide(request: Request) -> Verdict[Outcome, Outcome] to fn decide(request: Request) -> Outcome?.
I'm yet to make the const-ification thing happen, which will slightly tilt the diffstat towards additions, but hopefully not by much.
1+ more replies (not shown)
bonfire.cafe
A space for Bonfire maintainers and contributors to communicate
Automatic federation enabled