Claude
4def198cc0
- CLAUDE.md、testing.md、publishing.md 的 .gitea 路径改为实际的 .github/workflows； 说明 verify.yml 在 PR 上运行 check - testing.md 中带日期的逐轮结果迁到 docs/records/ - README 精简为简介、快速开始和入口；配置与迁移细节只保留在 operations.md - channel-messaging-v2.md 更名为 channel-messaging-contract.md，与 transfer-coordination.md 一起译为中文 - benchmarks/results 中两个记录改为日期前缀命名 - 新增 compileBenchmarks 并挂到 check，防止 benchmarks/ 静默失效 - .gitignore 删除不存在的 deployment/ 条目 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GGL9BudzWm4dUQgBsQw4Qr
2.9 KiB
Proxy session offline PLAY echo benchmark
Run date: 2026-09-25 (Asia/Shanghai)
Repository revision: 38b02266e6db7b1583116f53358cf97ab42f27b1
Runtime: Zulu OpenJDK 25.0.4.1, Windows PowerShell 7.6.5, AMD Ryzen 7 7800X3D
This measures only an offline synthetic Minecraft 1.7.10 PLAY packet echo. The direct case connects the benchmark client to the fake backend. The proxy case connects the same client to ProxySessionListener, which connects to that same backend. The backend sends Login Success, a valid Join Game and Position and Look, then echoes framed PLAY payloads. The client reads login and both initial PLAY packets and completes warmup before timing. The synthetic body starts with packet id 0x03 and fills the rest deterministically.
The two paths run in the same JVM with the same fake backend, client code, payload and per-phase connection concurrency. Phases alternate order across repeats. Login and warmup are excluded from timings. Throughput uses application payload bytes sent one way (excluding frame length bytes). Latency is measured from writing one request frame through reading its echo.
Small smoke
Command:
.\benchmarks\run-proxy-session.ps1 -Connections 1 -Messages 10 -Warmup 2 -Payload 64 -Repeats 1
| Phase | Measured roundtrips | Echoes correct | Roundtrips/s | One-way MiB/s | p50 ms | p95 ms | p99 ms |
|---|---|---|---|---|---|---|---|
| Direct | 10 | 10/10 | 10,377.8 | 0.63 | 0.071 | 0.122 | 0.122 |
| Proxy | 10 | 10/10 | 2,926.8 | 0.18 | 0.286 | 0.595 | 0.595 |
Two warmup echoes ran per phase and were also checked, then excluded from measured counts and timings.
Larger smoke
Command:
.\benchmarks\run-proxy-session.ps1 -Connections 4 -Messages 1000 -Warmup 100 -Payload 1024 -Repeats 2
Each phase measured 4,000 roundtrips; each phase also completed and checked 400 warmup echoes.
| Phase order | Measured roundtrips | Echoes correct | Roundtrips/s | One-way MiB/s | p50 ms | p95 ms | p99 ms |
|---|---|---|---|---|---|---|---|
| Direct 1 | 4,000 | 4,000/4,000 | 65,230.9 | 63.70 | 0.058 | 0.076 | 0.091 |
| Proxy 1 | 4,000 | 4,000/4,000 | 28,986.3 | 28.31 | 0.133 | 0.172 | 0.207 |
| Proxy 2 | 4,000 | 4,000/4,000 | 31,192.5 | 30.46 | 0.123 | 0.161 | 0.191 |
| Direct 2 | 4,000 | 4,000/4,000 | 62,558.6 | 61.09 | 0.062 | 0.079 | 0.093 |
Both commands completed :proxy-core:installDist, compilation and the benchmark successfully. Every measured response matched its request; each reported measured phase had zero echo errors.
These short loopback runs check benchmark correctness and provide an initial local comparison. They do not establish real client/server performance, Forge compatibility, pack behavior, network behavior outside loopback, or a stable capacity limit. The small sample has only ten measured roundtrips per phase; even the larger sample is a local microbenchmark and is sensitive to JVM and machine scheduling.