How? Enter Porffor
Porffor is my JS engine/runtime that compiles JavaScript ahead-of-time to WebAssembly and native binaries. What does that actually mean? You can compile JS files to tiny (<1MB), fast (millisecond-level) binaries:
~$ bat hi.js ─────┬────────────────────────────────────── 1 │ console.log ( "hello blog!" ) ─────┴────────────────────────────────────── ~$ porf native hi.js hi [ 271ms ] compiled hi.js -> hi ( 12.9KB ) ~$ du -h hi 16K hi ~$ ./hi hello blog!
Node and Bun offer “compile” options, but they bundle their runtime with your JS rather than actually compiling it as if it was C++ or Rust. Porffor does that, allowing for much smaller and faster binaries:
~$ deno compile -o hi_deno hi.js ~$ bun build --compile --outfile = hi_bun hi.js ~$ du -h hi* 16K hi 97M hi_bun 82M hi_deno 4.0K hi.js ~$ hyperfine -N "./hi" "./hi_deno" "./hi_bun" --warmup 5 Benchmark 1: ./hi Time ( mean ± σ ) : 631.4 µs ± 128.5 µs [ User: 294.5 µs, System: 253.1 µs ] Range ( min … max ) : 465.3 µs … 1701.3 µs 2762 runs Benchmark 2: ./hi_deno Time ( mean ± σ ) : 37.4 ms ± 1.7 ms [ User: 22.5 ms, System: 16.0 ms ] Range ( min … max ) : 33.8 ms … 42.2 ms 74 runs Benchmark 3: ./hi_bun Time ( mean ± σ ) : 15.9 ms ± 1.2 ms [ User: 8.7 ms, System: 9.6 ms ] Range ( min … max ) : 13.7 ms … 19.2 ms 175 runs Summary ./hi ran 25.24 ± 5.50 times faster than ./hi_bun 59.30 ± 12.36 times faster than ./hi_deno
What’s the trade-off? You have to re-invent the JS engine (and runtime) so it is still very early: limited JS support (but over 60% there) and currently no good I/O or Node compat (yet). But, we can use these tiny fast native binaries on Lambda!
Lambda
A few days ago I got Porffor running on Lambda, not simulated locally but really on AWS! I wrote a cold start benchmark for Node, LLRT (Amazon’s own experimental JS runtime optimizing cold starts) and Porffor running identical code:
export const handler = async () => { return { statusCode : 200 , headers : { "Content-Type" : "text/plain" }, body : "Hello from " + navigator . userAgent + " at " + Date () }; };
Since we’re benchmarking cold start, the workload does not matter as we are interested in just how we are running here (for context most Lambdas run for <1s, typically <500ms). I spent over a day just benchmarking and even with my biases, the results surprised me.
... continue reading