wehub-resource-sync
39 lines
1.3 KiB
Markdown
39 lines
1.3 KiB
Markdown
# benchmarking_nd4j
|
|
Simple Microbenchmarks for ND4J utilizing the JMH
|
|
|
|
Original credit to https://github.com/treo/benchmarking_nd4j
|
|
|
|
## Building
|
|
|
|
mvn clean package
|
|
|
|
## Running
|
|
|
|
java -jar target/benchmarks.jar
|
|
|
|
|
|
For a quick run, e.g. just the same Matrix Size progression as used in the Neanderthal benchmarks, run like this:
|
|
|
|
java -jar target/benchmarks.jar -f2 -i10 -wi 2 Neanderthal
|
|
|
|
## Running the memory profiler to detect memory leaks
|
|
Ensure jemalloc is installed with the profiler enabled. Something like:
|
|
wget https://github.com/jemalloc/jemalloc/releases/download/5.2.0/jemalloc-5.2.0.tar.bz2 && \
|
|
tar -xvf jemalloc-5.2.0.tar.bz2 && \
|
|
cd jemalloc-5.2.0 && \
|
|
./configure --enable-prof && \
|
|
make && \
|
|
make install
|
|
|
|
Run memory-prof.sh
|
|
|
|
This will run the above build process and the [MemoryPressureTest](src/main/java/org/nd4j/benchmark/memory/MemoryPressureTest.java) which will allocate a large amount of memory and then free it. The jemalloc profiler will be used to detect memory leaks.
|
|
|
|
## Choosing a BLAS Library
|
|
|
|
Since ND4J supports multiple blas libraries, you have to specify which one you actually want to use. For my own benchmarks I've been using MKL.
|
|
|
|
-Dorg.bytedeco.javacpp.openblas.load=mkl_rt
|
|
|
|
For more information see https://github.com/bytedeco/javacpp-presets/tree/master/openblas
|