Trimmed down cache performance comparison

[ljacomet]

Create a trimmed down version of Ehcache that does not support CacheLoaderWriter and then compare its performance with the current Ehcache class.

The goal is to have a measure of the cost of using functions and Store.compute* functions for all operations when there is no CacheLoaderWriter present.

[ljacomet]

Created a jmh based micro benchmark that compares Ehcache 3 with Ehcache 2.9.1.

Result on my machine with Java 1.8.0_31:

$> java -Xms2g -Xmx2g -jar target/benchmarks.jar -t 4 -f 1
(...)
# Run complete. Total time: 00:02:42

Benchmark                               Mode  Cnt          Score         Error  Units
CacheBenchmark.testComputeIfAbsentCHM  thrpt   20   80689508.131 ±  796255.458  ops/s
CacheBenchmark.testGetCHM              thrpt   20  329161816.150 ± 8217034.060  ops/s
CacheBenchmark.testGetEhcache2         thrpt   20   19483378.524 ±  235130.523  ops/s
CacheBenchmark.testGetEhcache3         thrpt   20   30813004.783 ± 1112848.209  ops/s

Will work on including results when running with a trimmed down version of Ehcache.get code path.

[ljacomet]

Test run with Java 1.6.0_65 (and without CHM.computeIfAbsent test):

$> java -Xms2g -Xmx2g -jar target/benchmarks.jar -t 4 -f 1
(...)
# Run complete. Total time: 00:02:03

Benchmark                        Mode  Cnt          Score         Error  Units
CacheBenchmark.testGetCHM       thrpt   20  105037787.070 ? 2861028.627  ops/s
CacheBenchmark.testGetEhcache2  thrpt   20   16542517.765 ?  427401.347  ops/s
CacheBenchmark.testGetEhcache3  thrpt   20   25432309.954 ?  918279.696  ops/s

[ljacomet]

Created a Google doc spreadsheet of which the graph above is an extract. The document contains JMH test results and information on the machines it was run on.
Tests were run based on code as found on https://github.com/ljacomet/ehcache3/tree/issue-309. The last commit creates the trimmed down version of Cache.get path, the parent commit is what was used in normal tests.

According to my findings, there is currently not enough value in splitting the code to have a simpler cache un-aware of CacheLoaderWriter as there is no clear performance gain.