Skip to content

docs(memory): JMH benchmark baseline 2026-08-22 (SKEEP-003 #1005 follow-up) #1052

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
michalharakal merged 1 commit into from
Aug 23, 2026
Merged

docs(memory): JMH benchmark baseline 2026-08-22 (SKEEP-003 #1005 follow-up) #1052

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
38 changes: 38 additions & 0 deletions docs/design/memory/baseline-2026-08-22.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,38 @@
# Benchmark baseline — 2026-08-22 (SKEEP-003 S0.9)

Reference numbers recorded **before** the memory-architecture migration (M1 `Storage`/`TensorView` façades, registry dispatch), to compare against in later slices (decision #6 budget: matmul within noise, elementwise ≤ 3 %).

- Commit: `5611b918` (develop + golden parity gate)
- Machine: Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz · Linux-6.8.0-60-generic-x86_64-with-glibc2.34 · openjdk version "25.0.3" 2026-04-21
- Suite: `./gradlew :skainet-backends:benchmarks:jvm-cpu-jmh:jmh` (JMH, fork 1, 3 warmup, 5 measurement iterations, `--enable-preview --add-modules jdk.incubator.vector`), run with nothing else building on the machine.
- Not included: `:skainet-lang:skainet-lang-core:jvmBenchmark` (kotlinx-benchmark `StorageBenchmarks`/`TensorBenchmarks`/`TurboQuantBenchmarks`) — with the module's default iteration settings a full run takes hours; run it filtered per class when a slice touches the storage hot paths and attach the numbers to that PR.

| Benchmark | Params | Mode | Cnt | Score | ± | Units |
|---|---|---|---|---:|---:|---|
| `add_1M_fp32` <br><sub>ElementwiseAdd1MBench</sub> | vectorEnabled=true | thrpt | 5 | 1.705 | 0.030 | ops/ms |
| `add_1M_fp32` <br><sub>ElementwiseAdd1MBench</sub> | vectorEnabled=false | thrpt | 5 | 1.217 | 0.011 | ops/ms |
| `mean_1M_fp32` <br><sub>Reductions1MBench</sub> | vectorEnabled=true | thrpt | 5 | 1.026 | 0.011 | ops/ms |
| `mean_1M_fp32` <br><sub>Reductions1MBench</sub> | vectorEnabled=false | thrpt | 5 | 1.026 | 0.004 | ops/ms |
| `sum_1M_fp32` <br><sub>Reductions1MBench</sub> | vectorEnabled=true | thrpt | 5 | 1.026 | 0.003 | ops/ms |
| `sum_1M_fp32` <br><sub>Reductions1MBench</sub> | vectorEnabled=false | thrpt | 5 | 1.027 | 0.006 | ops/ms |
| `matmul_fp32_square` <br><sub>KernelMatmulBench</sub> | provider=scalar, size=256 | avgt | 5 | 22.858 | 1.734 | ms/op |
| `matmul_fp32_square` <br><sub>KernelMatmulBench</sub> | provider=scalar, size=512 | avgt | 5 | 207.298 | 0.961 | ms/op |
| `matmul_fp32_square` <br><sub>KernelMatmulBench</sub> | provider=scalar, size=1024 | avgt | 5 | 1685.182 | 88.376 | ms/op |
| `matmul_fp32_square` <br><sub>KernelMatmulBench</sub> | provider=panama, size=256 | avgt | 5 | 1.186 | 0.008 | ms/op |
| `matmul_fp32_square` <br><sub>KernelMatmulBench</sub> | provider=panama, size=512 | avgt | 5 | 9.157 | 0.227 | ms/op |
| `matmul_fp32_square` <br><sub>KernelMatmulBench</sub> | provider=panama, size=1024 | avgt | 5 | 80.878 | 0.374 | ms/op |
| `matmul_fp32_square` <br><sub>MatmulBench</sub> | blasEnabled=true, size=256, vectorEnabled=true | avgt | 5 | 1.226 | 0.010 | ms/op |
| `matmul_fp32_square` <br><sub>MatmulBench</sub> | blasEnabled=true, size=256, vectorEnabled=false | avgt | 5 | 21.129 | 1.062 | ms/op |
| `matmul_fp32_square` <br><sub>MatmulBench</sub> | blasEnabled=true, size=512, vectorEnabled=true | avgt | 5 | 9.216 | 0.094 | ms/op |
| `matmul_fp32_square` <br><sub>MatmulBench</sub> | blasEnabled=true, size=512, vectorEnabled=false | avgt | 5 | 209.844 | 0.821 | ms/op |
| `matmul_fp32_square` <br><sub>MatmulBench</sub> | blasEnabled=true, size=1024, vectorEnabled=true | avgt | 5 | 82.036 | 2.492 | ms/op |
| `matmul_fp32_square` <br><sub>MatmulBench</sub> | blasEnabled=true, size=1024, vectorEnabled=false | avgt | 5 | 1709.814 | 5.381 | ms/op |
| `matmul_fp32_square` <br><sub>MatmulBench</sub> | blasEnabled=false, size=256, vectorEnabled=true | avgt | 5 | 1.225 | 0.043 | ms/op |
| `matmul_fp32_square` <br><sub>MatmulBench</sub> | blasEnabled=false, size=256, vectorEnabled=false | avgt | 5 | 20.479 | 0.784 | ms/op |
| `matmul_fp32_square` <br><sub>MatmulBench</sub> | blasEnabled=false, size=512, vectorEnabled=true | avgt | 5 | 9.371 | 0.162 | ms/op |
| `matmul_fp32_square` <br><sub>MatmulBench</sub> | blasEnabled=false, size=512, vectorEnabled=false | avgt | 5 | 199.811 | 46.629 | ms/op |
| `matmul_fp32_square` <br><sub>MatmulBench</sub> | blasEnabled=false, size=1024, vectorEnabled=true | avgt | 5 | 82.912 | 0.778 | ms/op |
| `matmul_fp32_square` <br><sub>MatmulBench</sub> | blasEnabled=false, size=1024, vectorEnabled=false | avgt | 5 | 1757.350 | 126.247 | ms/op |
| `matmul_q4k_panama` <br><sub>QuantizedMatmulBench</sub> | shape=1024-1024 | avgt | 5 | 0.112 | 0.017 | ms/op |
| `matmul_q4k_panama` <br><sub>QuantizedMatmulBench</sub> | shape=4096-1024 | avgt | 5 | 0.336 | 0.003 | ms/op |
| `matmul_q4k_panama` <br><sub>QuantizedMatmulBench</sub> | shape=4096-4096 | avgt | 5 | 1.223 | 0.017 | ms/op |
Loading