biboudis / lightweight-streams Goto Github PK

This library employs the same pattern that the standard Java 8 library does but without iterators/spliterators and is based on lambdas/continuation-passing only. The programming model consists of combinators that wrap a source with a datatype which implements the LStream interface. Any combinators that are applied, effectively declare the transformations needed to be applied when an eager combinator is met.

source - lazy - lazy - lazy - eager

As in Java, boxing is avoided with hand specialized combinators (e.g. LongLStream).

The purpose of this library is to make the study of the lambda inlining decisions that the JVM makes easier.

# run unit tests
mvn -q test

# build benchmarks über-jar
mvn clean package -Dskiptests

# run benchmarks
java -Xmx2g -Xms2g -XX:-TieredCompilation \
     -Dbenchmark.N=1000000 -Dbenchmark.F=3000000 -Dbenchmark.N_outer=10000 -Dbenchmark.N_inner=10 \
     -jar target/microbenchmarks.jar -wi 5 -i 5 -f 1 -gc true -tu ms ".*"

in-progress

In C2, with large MaxInlineLevel, LStreams are slightly better in the cartesian product case, they are considerably slower in the sort case due to the sorting algorithm and in all other cases, numbers are similar.

In Graal, all megamorphic boxed/primitive LStreams with one lambda run 50-100% faster. Nearly all megamorphic primitive LStreams run considerably faster!

You can take a look at the most current measurements, that include baseline, LStreams and Java 8 Streams in the measurements files for both C2 and Graal:

• C2 benchmarks
• Graal benchmarks
• and in spreadsheet

• Inspired by our work on Clash of the Lambdas
• Nessos/Streams in F#
• sml-streams in Standard ML on MLton
• Graal
• j.u.stream code / docs

• JMH for microbenchmarking
• junit
• Guava