Add RecyclerBytesStreamOutput writeVInt benchmark

[mhl-b]

A different view on RecyclerBytesStreamOutput write benchmark. Existing RecyclerBytesStreamOutputBenchmark measures complex randomized input covering multiple paths. I think there is value in measuring individual hot-paths with simple inputs, especially in case of VInt there is an optimization for 1-byte int.

Differences from RecyclerBytesStreamOutputBenchmark:

• Measure throughput, not average time. RecyclerBytesStreamOutput has page recycling amortized cost that impacts individual writes. Throughput-mode smoothes these spikes.
• Constant folding. Despite randomized input, this input stays the same on every iteration of RecyclerBytesStreamOutput. I use constant values with blackholes.

Currently covers only VInt's but easy to extend for other methods.

Example run

./gradlew -p benchmarks run --args 'RecyclerBytesStreamOutputWriteBenchmark'

Benchmark                                            Mode  Cnt   Score   Error   Units
RecyclerBytesStreamOutputWriteBenchmark.writeVInt1  thrpt    5  77.812 ± 0.759  ops/ms
RecyclerBytesStreamOutputWriteBenchmark.writeVInt2  thrpt    5  61.969 ± 4.399  ops/ms
RecyclerBytesStreamOutputWriteBenchmark.writeVInt3  thrpt    5  45.292 ± 3.572  ops/ms
RecyclerBytesStreamOutputWriteBenchmark.writeVInt4  thrpt    5  32.945 ± 1.035  ops/ms

[elasticsearchmachine]

Pinging @elastic/es-distributed-coordination (Team:Distributed Coordination)

[DaveCTurner]

For my education, what is the difference here? Is it more than just turning the fraction over so it reports the average of operations-per-unit-time vs time-per-unit operation?

[mhl-b]

I believe it's "just" turning fraction over. https://javadoc.io/doc/org.openjdk.jmh/jmh-core/1.29/org/openjdk/jmh/annotations/Mode.html#Throughput.

public static final Throughput
Throughput: operations per unit of time.
Runs by continuously calling Benchmark methods, counting the total throughput over all worker threads. This mode is time-based, and it will run until the iteration time expires.

public static final AverageTime
Average time: average time per per operation.
Runs by continuously calling Benchmark methods, counting the average time to call over all worker threads. This is the inverse of Throughput, but with different aggregation policy. This mode is time-based, and it will run until the iteration time expires.

[DaveCTurner]

Does it need to be fixed or could it be any value from 1 to 127? Should be the same code anyway?

[mhl-b]

Can be any from 1 to 127. As long it's not final JMH should be able properly blackhole this property.

[DaveCTurner]

vint shifts by 7 bits, not 8, and there's a 5-byte case to consider too (either very large or negative)

[mhl-b]

I keep forgetting