OpenRefine Benchmarks (JMH)

This module contains Java Microbenchmark Harness (JMH) benchmarks for OpenRefine.

Prerequisites

• Java (JDK 11+)
• Maven

Build The Benchmark JAR

From the repository root:

mvn -pl benchmark -am -DskipTests package

This produces:

• benchmark/target/openrefine-benchmarks.jar

List Available Benchmarks

java -jar benchmark/target/openrefine-benchmarks.jar -l

General JMH Controls

Useful optional flags:

• -f <n>: forks
• -wi <n>: warmup iterations
• -i <n>: measurement iterations
• -w <duration>: warmup time per iteration (for example 200ms)
• -r <duration>: measurement time per iteration (for example 300ms)
• -t <n>: threads
• -p name=value: benchmark parameter override

Benchmarks In This Module

ToNumberBenchmark

Class:

• org.openrefine.benchmark.ToNumberBenchmark

Methods:

• toDoubleNew
• toLongNew

Built-in parameter:

• iterations: 1000, 10000

Examples:

# Run the whole class
java -jar benchmark/target/openrefine-benchmarks.jar ToNumberBenchmark

# Run one method
java -jar benchmark/target/openrefine-benchmarks.jar ToNumberBenchmark.toLongNew

# Override parameter
java -jar benchmark/target/openrefine-benchmarks.jar ToNumberBenchmark -p iterations=10000

ApacheLevenshteinBenchmark (distance-only comparison)

Class:

• org.openrefine.benchmark.ApacheLevenshteinBenchmark

Methods:

• apacheLevenshtein
• vicinoLevenshtein

Purpose:

• Compare raw Levenshtein distance computation performance between Apache Commons Text and Vicino.

Dataset selection order:

1. JVM property openrefine.benchmark.dataset (if provided)
2. main/tests/data/acm_large.txt (if present)
3. main/tests/data/government_contracts.csv
4. Synthetic fallback dataset

Examples:

# Run full distance comparison
java -jar benchmark/target/openrefine-benchmarks.jar ApacheLevenshteinBenchmark

# Run with explicit dataset path
java -Dopenrefine.benchmark.dataset=/root/data/itunes_amazon_tableB_large.txt \
  -jar benchmark/target/openrefine-benchmarks.jar ApacheLevenshteinBenchmark

# Run one method only
java -jar benchmark/target/openrefine-benchmarks.jar ApacheLevenshteinBenchmark.apacheLevenshtein

KNNLevenshteinClusteringBenchmark (end-to-end clustering)

Class:

• org.openrefine.benchmark.KNNLevenshteinClusteringBenchmark

Methods:

• apacheKNNClustering
• vicinoKNNClustering

Purpose:

• Compare full kNN clustering runtime, not just distance calls.

Built-in parameter (safe defaults):

• rowCount: 200, 500

Default warmup/measurement (in code):

• Warmup: 1 iteration x 200ms
• Measurement: 2 iterations x 300ms

Examples:

# Run end-to-end comparison with defaults
java -jar benchmark/target/openrefine-benchmarks.jar KNNLevenshteinClusteringBenchmark

# Run with explicit dataset and small row count
java -Dopenrefine.benchmark.dataset=/root/data/itunes_amazon_tableB_large.txt \
  -jar benchmark/target/openrefine-benchmarks.jar KNNLevenshteinClusteringBenchmark -p rowCount=200

# Extra-safe quick profile
java -jar benchmark/target/openrefine-benchmarks.jar KNNLevenshteinClusteringBenchmark \
  -f 1 -wi 0 -i 1 -r 1s -w 0s -p rowCount=200 -t 1

Notes On Runtime And Thermals

• End-to-end clustering benchmarks can be CPU-intensive.
• If your machine runs hot, use:
  • lower rowCount
  • -t 1
  • fewer iterations (-wi 0 -i 1) for quick checks
• JMH can appear to stay on one iteration for a while whenever each operation is CPU intensive.