Benchmarks
Accuracy (synthetic recovery)
Synthetic events with uniform full-focal-sphere coverage: without polarity errors CNCHASH recovers 10/10 known mechanisms (median Kagan rotation ~8 deg, the dang=5 grid discretization floor); with ~10% flipped polarities 9/10 remain within 25 deg. These checks are part of the test suite (tests/test_accuracy.py).
Note: Strike differences (40-80 deg) are normal - focal mechanisms have two orthogonal nodal planes that both satisfy polarity data.
Historical notebook results (HASH_Tests.ipynb): CNCHASH synthetic trials 300/300 successful solutions; HASH v1.2 60/60 (direct-runs).
Native backend measurements
Run locally with:
python benchmarks/benchmark_backends.py --events 1000 --seconds 2
Measured on a development machine (x86_64, gfortran). Values are for
30-station events with 30 MC trials (dang=5).
Thread scaling (portable build, 30 stations)
Threads |
ms/event |
Speedup |
Efficiency |
|---|---|---|---|
1 |
49.5 |
1.00x |
100% |
2 |
33.0 |
1.50x |
75% |
4 |
19.4 |
2.56x |
64% |
8 |
15.5 |
3.19x |
40% |
16 |
14.0 |
3.53x |
22% |
The native build (-DCNCHASH_NATIVE_MARCH=ON, AVX2) is roughly 2x
faster: ~23 ms/event at 1 thread and ~4 ms/event at 16 threads.
Config comparison (portable build, ms/event)
Config |
1 thread |
4 threads |
|---|---|---|
small (12 st) |
39 |
25 |
medium (30 st) |
49 |
20 |
large (80 st) |
119 |
41 |
dense (dang=2) |
49 |
19 |
nmc=100 |
156 |
63 |
Weakly constrained events (few stations, multiple solutions) are dominated by the MECH_PROB clustering, which now matches the original HASH working-set behavior; the parallel path still helps there.
Comparison with the original HASH v1.2 (same machine)
Implementation |
ms/event |
|---|---|
Original HASH v1.2 (1 thread) |
~145 |
CNCHASH portable (1 thread) |
~49 |
CNCHASH portable (4 threads) |
~20 |
CNCHASH native AVX2 (4 thr.) |
~8 |
Batch mode (medium config, portable build)
Threads |
Batch ev/s |
Scalar ev/s |
Batch speedup |
|---|---|---|---|
1 |
18.6 |
18.7 |
1.00x |
4 |
56.9 |
42.6 |
1.34x |
Numbers are machine-specific. Always re-run the benchmark on the target hardware before publishing performance claims.