A simple test of collection libraries on JVM.
Tested:
- Java 8 streams
- Kotlin sequence
- Kotlin list
- CopyList (simple functions which take a list and return a list)
- Imperative code
- Solid library
- Pass (my new streaming library prototype)
According to the benchmark, imperative code is the fastest. Not a big surprise.
The big surprise is the overall poor performance of lazy libraries. In most cases it is not better than CopyList - my trivial implementation of operations on list which allocated a new list every time.
The most interesting and perspective is Pass - my new look on collection libraries.
Instead of being lazy, it is strict, so it does not need to keep excessive temporary variables.
It also does not need to call hasNext for each processed item so it is thread-safe.
Pass adds from 2x to 5x less overhead than lazy streaming libraries.
My overall conclusion is that lazy evaluations are overcomplicated, slow, unsafe and overvalued.
JMH tested results on my Mac with max 3% error:
LIST SIZE 2 8 32 128 1024
BENCHMARK, nanoseconds
BenchmarkJmh.imperative 46 198 1158 4848 39345
BenchmarkJmh.pass 73 289 1424 5634 46695
BenchmarkJmh.passUnoptimized 87 342 1693 6741 54779
BenchmarkJmh.solid 109 472 2093 8626 71177
BenchmarkJmh.copyList 116 411 2074 8022 63981
BenchmarkJmh.kotlinSequence 116 471 2187 8945 71930
BenchmarkJmh.kotlinList 131 506 2189 8530 71752
BenchmarkJmh.java8 165 520 2153 8043 64735
OVERHEAD, nanoseconds
BenchmarkJmh.imperative 0 0 0 0 0
BenchmarkJmh.pass 27 91 266 785 7350
BenchmarkJmh.passUnoptimized 41 144 535 1892 15433
BenchmarkJmh.solid 63 274 935 3777 31831
BenchmarkJmh.copyList 69 213 916 3173 24635
BenchmarkJmh.kotlinSequence 70 274 1029 4097 32584
BenchmarkJmh.kotlinList 85 308 1031 3682 32407
BenchmarkJmh.java8 119 323 996 3195 25390