Skip to content

Latest commit

 

History

History
20 lines (14 loc) · 938 Bytes

BENCH.md

File metadata and controls

20 lines (14 loc) · 938 Bytes

Micro-benchmarks

The benchmarks are done with JMH framework.

Execute make bench to run them.

The predictions are performed with the file generated.model (126 features and 1000 trees) that is generated by Xgboost4jCompatibilySpec unit test.

The figures below were done with a single thread on a Intel(R) Core(TM) i5-6400

Benchmark                           Mode  Cnt   Score    Error   Units
BenchPredict.predictContrib        thrpt    5   6.300 ±  0.006  ops/ms
BenchPredict.predictContribApprox  thrpt    5  87.402 ± 11.126  ops/ms
BenchPredict.predict               thrpt    5  98.762 ±  1.583  ops/ms
BenchXgboost4j.predict             thrpt    5  68.195 ±  4.002  ops/ms
BenchXgboost4j.predictContrib      thrpt    5   7.269 ±  1.721  ops/ms

Note that Xgboost4j figures are really fluctuent and seem to depend a lot upon system status (because of JNI?)