سرعت پایین مدل Full Tunnel #1038

Activity

1. therealaleph commented on May 10, 2026

   

   therealaleph

   on May 10, 2026

   Owner

   More actions

   Reviewed via Anthropic Claude.

   @mirbaagheri — your analysis (and ChatGPT's) is partially right, partially missing context:

   Correctly identified bottlenecks:

   • Base64 overhead ✅ — 33% bandwidth bloat, real cost on hot paths.
   • JSON encode/decode per batch ✅ — Apps Script's V8 isolate is slow at this.
   • UrlFetchApp call overhead ✅ — Google's edge adds 100-300ms per call regardless of payload.
   • HTTP POST per packet ✅ — was true in v1.7 era.

   What's already addressed in code (you can skip the upstream-architectural-rewrite suggestion):

   1. "HTTP POST per stream chunk instead of per packet" — already done. Look at coalesce_step_ms / coalesce_max_ms in config.json. Default packs up to 1000ms of incoming ops into one batch (50 ops max). Increase coalesce_max_ms: 5000 if you want larger chunks at the cost of latency.

   2. "Skip GAS for Full mode" — that's the entire point of Full mode + a VPS with cloudflared / ngrok / domain. You're routing client → Apps Script → tunnel-node, but the tunnel-node is the part that talks to internet directly. The Apps Script middleman exists because direct Iran → VPS over plain HTTPS is what gets DPI'd; Apps Script edge is what bypasses DPI. Without it, the whole project's value goes away.

   3. "Base64 removal" — possible but invasive. Apps Script's UrlFetchApp requires response bodies as text or binary blob; pure-binary tunneling without base64 means changing the JSON envelope to a different wire format (e.g. MessagePack or raw multipart). PR feat: zero-copy full-tunnel mux + base64 off mux thread #881 (already merged, v1.9.18) moved base64 encoding off the single mux thread as a halfway measure — saves CPU time, not bytes. Full base64 removal would need a wire-protocol break.

   4. "Large chunk multiplexing" — PR feat: per-session pipelining via seq-ordered data ops #903 (in community testing) is exactly this: per-session pipelining with seq-ordered data ops, up to 2 in-flight per session. Track that PR.

   5. "H2 vs H1 efficiency" — PR perf: skip H2 for full-tunnel batch requests #1040 (just opened) skips H2 for tunnel-batch requests (where it has nothing to multiplex anyway) while keeping H2 for relay mode. Concrete A/B data: 0/30 vs 8-10/30 long-poll stalls.

   v1.9.20 (just shipped) fixed the warm-up race that was costing many users 30s cold-start timeouts. Update first; baseline perf today is meaningfully better than 1 week ago.

   For your specific setup (own VPS + SSL, no Cloudflare): you're already on the best Full-mode path. Speed bottleneck remaining is:

   • Apps Script's UrlFetchApp 100-300ms RTT per batch (architectural, not bug)
   • Plus VPS network latency

   For "fast at any cost": don't use mhrv-rs Full mode. Run xray-core with Reality protocol directly on the same VPS. mhrv-rs's value is DPI-bypass + free Apps Script quota; pure speed is a different transport.

   می‌بندم — perf work is ongoing (#903 + #1040 active). Update to v1.9.20 + try coalesce_max_ms: 3000 and report back if speed improved.
2. therealaleph

   closed this as completedon May 10, 2026