Skip to content

Latest commit

 

History

History
159 lines (135 loc) · 10 KB

benchmark-windows.md

File metadata and controls

159 lines (135 loc) · 10 KB

Windows IPC Benchmark Results

Benchmark of the cross-language Windows IPC transports across the full directed matrix:

  • c-native
  • c-msys
  • rust-native
  • go-native

Direction matters here. client=A -> server=B is measured separately from client=B -> server=A.

Environment

  • OS: Windows 11
  • Host type: Hyper-V VM
  • Duration: 5 seconds per timed run
  • Protocol: single-connection request/response increment loop
  • C native build: MinGW-w64 native Windows binary
  • C MSYS build: MSYS runtime binary linked against msys-2.0.dll
  • Rust build: native Windows binary
  • Go build: native Windows binary

Transport Profiles

Profile Description
Named Pipe CreateNamedPipeW / ReadFile / WriteFile kernel IPC
SHM HYBRID Shared memory ring with spin plus kernel event fallback, negotiated over the initial Named Pipe handshake

Results

Maximum throughput, SHM HYBRID

Pair Method Throughput (rps) p50 (us) p95 (us) p99 (us) Client CPU Server CPU Total CPU
c-native->c-native c-shm-hybrid 3279438.39 0.20 0.26 0.28 0.894 0.769 1.663
c-native->c-msys c-shm-hybrid 3259295.13 0.23 0.29 0.33 0.903 0.799 1.702
c-native->rust-native rust-shm-hybrid 3210545.77 0.22 0.29 0.31 0.850 0.829 1.679
c-native->go-native go-shm-hybrid 2046836.10 0.00 0.00 0.00 1.200 0.882 2.082
c-msys->c-native c-shm-hybrid 3375322.22 0.21 0.26 0.29 0.925 0.756 1.681
c-msys->c-msys c-shm-hybrid 3409185.94 0.22 0.27 0.30 0.906 0.826 1.732
c-msys->rust-native rust-shm-hybrid 3332492.49 0.20 0.27 0.30 0.950 0.842 1.792
c-msys->go-native go-shm-hybrid 2065578.26 0.00 0.00 0.00 1.206 0.876 2.082
rust-native->c-native c-shm-hybrid 3196945.33 0.22 0.27 0.30 0.891 0.758 1.649
rust-native->c-msys c-shm-hybrid 3148056.03 0.24 0.28 0.32 0.900 0.802 1.702
rust-native->rust-native rust-shm-hybrid 3344758.28 0.21 0.27 0.29 0.894 0.838 1.732
rust-native->go-native go-shm-hybrid 2047591.61 0.00 0.00 0.00 1.159 0.857 2.016
go-native->c-native c-shm-hybrid 1590707.25 0.35 0.48 7.18 0.916 1.094 2.010
go-native->c-msys c-shm-hybrid 1667772.62 0.31 0.41 6.51 0.869 1.058 1.927
go-native->rust-native rust-shm-hybrid 1647485.91 0.32 0.41 6.45 0.887 1.169 2.056
go-native->go-native go-shm-hybrid 1093866.75 0.00 0.00 0.00 1.231 1.158 2.389

Maximum throughput, Named Pipe

Pair Method Throughput (rps) p50 (us) p95 (us) p99 (us) Client CPU Server CPU Total CPU
c-native->c-native c-npipe 17604.37 39.08 3742.59 18560.49 0.212 0.190 0.402
c-native->c-msys c-npipe 18051.11 32.74 3037.73 15262.83 0.235 0.232 0.467
c-native->rust-native rust-named-pipe 18418.94 39.83 3756.52 15068.80 0.259 0.221 0.480
c-native->go-native go-named-pipe 16811.07 0.00 500.60 511.30 0.297 0.216 0.513
c-msys->c-native c-npipe 17969.51 39.14 3695.47 14829.90 0.221 0.246 0.467
c-msys->c-msys c-npipe 17172.68 41.78 4153.14 15806.03 0.244 0.188 0.432
c-msys->rust-native rust-named-pipe 18336.34 39.21 3689.14 14787.17 0.202 0.218 0.420
c-msys->go-native go-named-pipe 15153.88 0.00 501.80 517.90 0.325 0.221 0.546
rust-native->c-native c-npipe 18138.95 41.04 3882.04 19255.78 0.264 N/A N/A
rust-native->c-msys c-npipe 17296.77 38.44 7327.25 18483.16 0.233 N/A N/A
rust-native->rust-native rust-named-pipe 17621.21 37.91 3598.02 14551.40 0.284 N/A N/A
rust-native->go-native go-named-pipe 18182.87 0.00 500.30 511.60 0.337 N/A N/A
go-native->c-native c-npipe 18424.69 38.31 7175.56 17896.40 0.259 0.318 0.577
go-native->c-msys c-npipe 18603.50 39.27 7300.46 18339.88 0.246 0.318 0.564
go-native->rust-native rust-named-pipe 16048.27 40.06 11052.28 22003.45 0.201 0.318 0.519
go-native->go-native go-named-pipe 18161.59 0.00 500.20 509.60 0.359 0.333 0.692

Rate-limited at 100k rps, SHM HYBRID

Pair Method Throughput (rps) p50 (us) p95 (us) p99 (us) Client CPU Server CPU Total CPU
c-native->c-native c-shm-hybrid 99842.34 0.28 0.36 26.97 0.921 0.918 1.839
c-native->c-msys c-shm-hybrid 99998.05 0.24 0.31 24.45 0.930 0.917 1.847
c-native->rust-native rust-shm-hybrid 99995.00 0.26 0.35 27.50 0.549 0.543 1.092
c-native->go-native go-shm-hybrid 99990.75 0.00 0.00 0.00 0.228 0.159 0.387
c-msys->c-native c-shm-hybrid 99783.79 0.26 0.33 27.11 0.880 0.866 1.746
c-msys->c-msys c-shm-hybrid 100000.07 0.24 0.33 32.40 0.864 0.867 1.731
c-msys->rust-native rust-shm-hybrid 99996.54 0.03 0.04 2.91 0.599 0.561 1.160
c-msys->go-native go-shm-hybrid 99990.45 0.00 0.00 0.00 0.228 0.178 0.406
rust-native->c-native c-shm-hybrid 100000.05 0.26 0.33 24.58 0.933 0.915 1.848
rust-native->c-msys c-shm-hybrid 100000.18 0.27 0.35 25.63 0.936 0.911 1.847
rust-native->rust-native rust-shm-hybrid 99959.98 0.03 0.04 2.94 0.614 0.616 1.230
rust-native->go-native go-shm-hybrid 99990.29 0.00 0.00 0.00 0.237 0.144 0.381
go-native->c-native c-shm-hybrid 100000.06 0.40 12.70 67.25 0.889 1.174 2.063
go-native->c-msys c-shm-hybrid 99999.30 0.38 11.81 62.84 0.864 1.099 1.963
go-native->rust-native rust-shm-hybrid 99990.86 0.38 12.04 52.92 0.736 1.015 1.751
go-native->go-native go-shm-hybrid 99990.49 0.00 0.00 0.00 0.325 0.352 0.677

Rate-limited at 10k rps, Named Pipe

Pair Method Throughput (rps) p50 (us) p95 (us) p99 (us) Client CPU Server CPU Total CPU
c-native->c-native c-npipe 10000.11 42.25 7437.44 18565.18 0.529 0.208 0.737
c-native->c-msys c-npipe 9998.91 40.02 7128.49 17762.96 0.489 0.124 0.613
c-native->rust-native rust-named-pipe 9962.93 42.47 7442.37 14872.91 0.199 0.142 0.341
c-native->go-native go-named-pipe 9999.33 0.00 456.90 507.70 0.262 0.178 0.440
c-msys->c-native c-npipe 9998.49 43.69 7478.05 21937.61 0.486 0.168 0.654
c-msys->c-msys c-npipe 10000.08 42.97 7966.41 19704.49 0.489 0.156 0.645
c-msys->rust-native rust-named-pipe 10000.03 41.33 7277.43 18222.23 0.224 0.131 0.355
c-msys->go-native go-named-pipe 9999.24 0.00 490.40 508.00 0.303 0.134 0.437
rust-native->c-native c-npipe 10000.07 42.78 7461.31 21943.83 0.592 N/A N/A
rust-native->c-msys c-npipe 9999.57 38.74 6488.23 16139.50 0.495 N/A N/A
rust-native->rust-native rust-named-pipe 9999.50 42.14 7557.97 18370.55 0.206 N/A N/A
rust-native->go-native go-named-pipe 9999.26 0.00 493.20 510.50 0.275 N/A N/A
go-native->c-native c-npipe 10000.11 43.01 11043.92 21672.94 0.517 0.177 0.694
go-native->c-msys c-npipe 9998.50 43.15 11215.60 21993.44 0.402 0.221 0.623
go-native->rust-native rust-named-pipe 9993.73 44.78 11219.74 21994.85 0.277 0.226 0.503
go-native->go-native go-named-pipe 9999.20 0.00 499.50 512.00 0.275 0.231 0.506

Key Findings

  • c-msys and c-native are effectively in the same performance class on the Windows transport.
    • Example, SHM max:
      • c-native->c-native: 3.28M rps
      • c-native->c-msys: 3.26M rps
      • c-msys->c-native: 3.38M rps
      • c-msys->c-msys: 3.41M rps
  • The transition requirement is therefore met in practice:
    • MSYS-built C can talk to native Rust and Go.
    • Native Rust and Go can talk back to both c-native and c-msys.
  • SHM HYBRID remains the only path that reaches multi-million request rates on Windows.
    • C and Rust clients driving C or Rust servers cluster around 3.1M to 3.4M rps.
    • Go clients are materially slower in SHM max-rate runs:
      • about 1.59M to 1.67M rps to C/Rust servers
      • about 1.09M rps for go-native->go-native
  • Named Pipe throughput stays in the same general band across the full matrix:
    • about 15k to 18.6k rps at max rate
    • about 10k rps when rate-limited
  • At 100k rps SHM, all directed pairs stay near the target rate, but client choice still affects tail latency and CPU.
    • Go clients show much higher p95/p99 tails against C/Rust SHM servers than C/Rust clients do.

Reproducing

# Native Windows helper build (C native + Rust + Go)
cmake -S . -B build-mingw -G "MinGW Makefiles" -DCMAKE_BUILD_TYPE=Release
cmake --build build-mingw

# MSYS-runtime C helper build
cmake -S . -B build-msys-ninja -G Ninja -DCMAKE_BUILD_TYPE=Release -DNETIPC_BUILD_HELPERS=OFF
cmake --build build-msys-ninja --target netipc-live-c

# Full 64-row matrix
bash tests/run-live-win-bench.sh

Known Limitations

  • Go latency columns currently print 0.00 for many SHM rows, so the useful signal there is throughput and CPU rather than the raw latency fields.
  • Rust client rows still miss some Named Pipe server CPU values (N/A) because the benchmark metadata is not always emitted before the server exits.
  • These numbers are from a Windows 11 Hyper-V VM. Relative differences inside this matrix are useful; absolute numbers should not be treated as bare-metal limits.