JMH Benchmarks: Avoid byte array copy for reads in NMA valueSerializer

[aaronriekenberg]

GCP VM Perf results with this change are below, compare with baseline: https://github.com/target/native_memory_allocator/wiki/Benchmark-results-in-GCP-VM

• NMA tps is 3.7x more in read_only test (257K -> 970K)
• NMA tps is 2.1x more in readwrite test (256K -> 555K)
• Results for RocksDB do not change.

~/jdk-17.0.6/bin/java -jar ./benchmarks-jmh.jar -wi 10 -i 20 -f 1 -gc true -foe true -tg 32 -jvmArgs '-Xmx4G -Drocksdb.dir=/tmp/tmpfs/rocksdb' read_only

Benchmark                         (cacheType)   Mode  Cnt       Score      Error  Units
OffHeapGetPutBenchmark.read_only          NMA  thrpt   20  970632.328 ± 8836.803  ops/s
OffHeapGetPutBenchmark.read_only      RocksDB  thrpt   20  120857.160 ± 1444.857  ops/s


~/jdk-17.0.6/bin/java -jar ./benchmarks-jmh.jar -wi 10 -i 20 -f 1 -gc true -foe true -tg 32 -jvmArgs '-Xmx4G -Drocksdb.dir=/tmp/tmpfs/rocksdb' write_only

Benchmark                          (cacheType)   Mode  Cnt       Score      Error  Units
OffHeapGetPutBenchmark.write_only          NMA  thrpt   20  199977.906 ± 6923.452  ops/s
OffHeapGetPutBenchmark.write_only      RocksDB  thrpt   20    2466.037 ±  492.823  ops/s


~/jdk-17.0.6/bin/java -jar ./benchmarks-jmh.jar -wi 10 -i 20 -f 1 -gc true -foe true -tg 24,8 -jvmArgs '-Xmx4G -Drocksdb.dir=/tmp/tmpfs/rocksdb'   readwrite

Benchmark                                       (cacheType)   Mode  Cnt       Score       Error  Units
OffHeapGetPutBenchmark.readwrite                        NMA  thrpt   20  555568.405 ± 30179.022  ops/s
OffHeapGetPutBenchmark.readwrite:readwrite_get          NMA  thrpt   20  481717.573 ± 26025.353  ops/s
OffHeapGetPutBenchmark.readwrite:readwrite_put          NMA  thrpt   20   73850.832 ±  4198.050  ops/s
OffHeapGetPutBenchmark.readwrite                    RocksDB  thrpt   20  109377.958 ± 17557.734  ops/s
OffHeapGetPutBenchmark.readwrite:readwrite_get      RocksDB  thrpt   20  107904.098 ± 17242.323  ops/s
OffHeapGetPutBenchmark.readwrite:readwrite_put      RocksDB  thrpt   20    1473.860 ±   429.475  ops/s

[aaronriekenberg]

toTrimmedArray makes a copy of the byte array on every read, asByteBuffer does not make a copy - this takes advantage of the thread-local read buffer in NativeMemoryMap

[kellar]

LGTM