HTTP/2 PROXY / OCTOBER 17, 2026

Independent API streams.

Five rounds on one AWS c7g.large, one worker, one verified TLS / HTTP/2 connection. Client and backend share the host over loopback. Sixteen API requests each wait 800 ms at the backend while fast requests use sibling streams. Each round checks ten fast API responses, ten static responses and all sixteen delayed responses. TCP_NODELAY is enabled in the client, server and backend. Timings include client and backend scheduling. This is a multiplexing regression, not a capacity test or a comparison against other web servers. CPU and RAM were not sampled.

Zero errors across 180 verified responses. All sixteen delayed requests finish in a median 805.62 ms per round. The backend still takes its original 800 ms; the other streams can finish while it waits. p95 uses the lower observed quantile; all fifty samples per probe are linked below.

Client configuration check

The initial trial had Nagle enabled in the test client and showed roughly 40 ms delays in both probes. A repeat with the identical server binary and only client TCP_NODELAY changed removed that delay. Initial samples and controlled client check are retained. This is a client configuration effect, not a server speedup. The final build also corrects HTTP/2 access-log protocol and client-address reporting.

Protocol and limits

Clients use HTTP/2 over TLS. Superweb opens one cleartext HTTP/1.1 backend connection per request. Uploads stream with or without Content-Length, up to 1 MiB. Raw upload queues grow to 65,535 bytes per stream; response queues default to 16 KiB. Cancellation and timeouts affect individual streams. Upstream HTTP/2, upstream TLS, gRPC trailer forwarding and WebSocket extended CONNECT are not implemented.

Validation

The normal server suites and ASan/UBSan passed on macOS ARM64 and Linux ARM64. Seven HTTP/2 integration tests cover concurrent API/static streams, uploads before END_STREAM, 16 MiB responses, flow-control pauses, cancellation, malformed framing, trailers, auth, cookies, timeouts, early responses and repeated disconnects with an eight-slot pool. HTTP/1.1 and WebSocket regressions also passed. The bundled Intel macOS build remains HTTP/1.1 only.

Measured build

Ubuntu 24.04 ARM64, -O3 -flto -mcpu=neoverse-n1. Linux binary SHA-256:
bc2e27d386813b9b2b7f1f689ede010fd6ca8da58c7923ef753e5519aeb27ad2

The prior release did not proxy HTTP/2, so this report makes no before/after latency claim. Earlier game-traffic and four-server static results retain their measured binaries; they have not been rerun for this release.

Every sample and environment · Benchmark harness · HTTP/2 fixture · HTTP backend fixture · Shared WebSocket fixture · Reproduction · Checksums