Benchmarks

Tulpar’s tagline is “as easy as Python, as fast as C”. The CPU side of that claim is verified on every commit; the HTTP side gets us into Node.js territory single-threaded.

CPU benchmarks

benchmarks/loopsum.tpr and benchmarks/fib.tpr use the typed AOT path (explicit : int return type) for native LLVM i64 codegen.

Benchmark	C (gcc -O2)	Tulpar AOT	Ratio
loopsum (10M sum)	67 ms	88 ms	1.31× C
fib(35) recursive	83 ms	114 ms	1.37× C

Tulpar AOT lands in the 1.3–1.4× C range — same neighbourhood as Rust and Go on identical hardware. Best-of-3 wall-clock times on Windows 11 + MinGW64.

HTTP throughput

benchmarks/http_bench.py — 5000 GET requests over 4 keep-alive TCP connections, single-threaded servers, JSON {"hello":"world"} body.

Server	Wall (s)	req/sec	Notes
Tulpar Wings	0.193	~26 000	Single-thread listen(), NODELAY on accept, dlsym handler cache.
Node.js http	0.184	~27 200	http.createServer, V8 22.x.
Python ThreadingHTTP	0.354	~14 100	ThreadingHTTPServer, single CPython process.

listen_async() enables a multi-threaded variant — handler dispatch still serialises under _wings_handler_mu until LLVM thread-local globals land, but parallel recv / send lifts throughput on keep-alive workloads where many connections sit idle.

What got us here

Hot-path optimisations applied:

• call(handler_name) dlsym cache (256-slot FNV-1a hash) — eliminates the symbol-table walk per request.
• TCP_NODELAY on accept — removes Nagle’s 40ms batching delay on small JSON responses (+13% req/sec).
• Static thread-local recv buffer — drops a 64 KB malloc/free pair per keep-alive request.
• Per-request arena reset — bounded memory on long-running servers without leaking.
• break / continue real codegen — was silently no-op’d before, prevents LLVM from emitting suboptimal phi nodes around induction variables.

Trade-offs we explicitly skipped (analysed, low value):

• Object-key inline caching (req["method"] ~0.3 % of HTTP path).
• String concat coalescing (a + b + c ~0.1 % of HTTP path).

Methodology

• Native compilers: gcc -O2. Rust: rustc -C opt-level=3. Go: go build defaults.
• HTTP results use single-threaded servers everywhere; running Node with cluster or Tulpar with a thread-pool mode is a separate measurement.
• Numbers regenerated via benchmarks/run_benchmarks.sh (CPU) and python benchmarks/http_bench.py (HTTP).