rmatch

Current Performance Comparison

Metric	rmatch	Java Regex	Ratio (rmatch/java)
5000 patterns	2.4s	4.1s	1.7x faster
Peak Memory	39MB	19MB	2.1x more memory
Pattern Loading	20MB	1MB	20.0x more memory
Matching Phase	5MB	4MB	1.2x more memory

Latest benchmark comparison between rmatch and native Java regex (java.util.regex.Pattern) on 5000 regex patterns against Wuthering Heights corpus. Updated: 2025-12-18 21:42 UTC

---

Performance Timeline Charts

rmatch Performance History

Java Regex Performance History

Performance Comparison (rmatch vs Java Regex)

Live performance tracking from macro benchmarks. Individual charts show execution time and memory usage patterns over time, while the comparison chart shows rmatch performance ratios relative to Java regex (values > 1.0 mean rmatch is slower/uses more memory).

---

rmatch

The project is getting closer to a state where it may be useful for others than myself, but it's not quite there yet. Be patient ;)

Key Performance Metrics

• Benchmark Data Sources: All performance data is sourced from benchmarks/results/
• JMH Microbenchmarks: Precise timing measurements with statistical confidence intervals
• Macro Benchmarks: End-to-end performance testing with real workloads
• Automated Tracking: Performance evolution tracked continuously via GitHub Actions

---

🚨 CRITICAL: Performance Validation Rule

"I will not merge anything to main that does not provably improve performance."

Development Guidelines for Performance Changes

ALL performance optimizations MUST:

1. ✅ Use Production-Scale Testing: Test with 5000+ regex workloads against real text corpora
2. ✅ Show Measurable Improvement: Demonstrate clear performance gains in comprehensive benchmarks
3. ❌ Never Trust Micro-Benchmarks: Small-scale synthetic tests are insufficient and often misleading
4. ❌ Never Trust Theoretical Improvements: Code that "should be faster" must prove it IS faster

Lessons Learned

• Enum switching theoretical 13.6% improvement → Actually slower in production
• Pattern matching instanceof → 1.7% performance regression
• Character classification optimizations → 2-9% performance regression

The Rule Exists Because: Brilliant optimization ideas are necessary, but they must be proven in our specific use case with realistic workloads before adoption.

---

GC Optimization for Java 25

rmatch includes tools to experiment with different Garbage Collector (GC) configurations on Java 25 to optimize memory usage and performance.

Quick Start

Validate GC configurations:

scripts/validate_gc_configs.sh

Run benchmarks with all GC variants:

make bench-gc-experiments

Available GC Configurations

• G1 (default): General-purpose GC with good throughput
• ZGC Generational: Low-latency GC for large heaps
• Shenandoah: Concurrent GC with predictable pause times
• Compact Object Headers: Reduces memory footprint (Java 25 feature)

Documentation

See GC_EXPERIMENTS.md for:

• Detailed usage instructions
• How to analyze results
• Applying findings to improve performance
• References to Java 25 GC improvements

The experiments help identify optimal GC settings for regex engines with high object churn, following recommendations from the JDK 25 Performance Improvements article.

---

⚡ Java 25 JIT Optimization (Production Ready)

rmatch includes proven JIT optimization techniques for Java 25 that provide significant performance improvements in production workloads.

Recommended Production Configuration

For optimal performance across all workload sizes:

export JAVA_OPTS="-Drmatch.engine=fastpath -Drmatch.prefilter=aho -XX:+TieredCompilation -XX:CompileThreshold=500"

Performance Results

Test Environment: Apple M2 Max (aarch64), macOS 26.0.1, OpenJDK 25 (Temurin-25+36-LTS)

Configuration	5K Patterns	10K Patterns
Baseline	10,230ms	21,656ms
Optimized	9,895ms (+3.3%)	19,297ms (+10.9%)

Performance characteristics may vary across different architectures. See FASTPATH_PERFORMANCE_ANALYSIS.md for detailed architecture specifications and cross-platform considerations.

Key Benefits

• ✅ Scales with workload size: 3.3% improvement at 5K, 10.9% at 10K patterns
• ✅ Consistent performance: 8% coefficient of variation across runs
• ✅ Production validated: Tested with comprehensive benchmarks using real text corpora
• ✅ Easy to apply: Simple environment variable configuration

Advanced Configuration

For specific workload tuning:

# Small workloads (≤5K patterns): Pure ASCII optimization
export JAVA_OPTS="-Drmatch.engine=fastpath -Drmatch.prefilter.threshold=99999 -XX:+TieredCompilation -XX:CompileThreshold=500"

# Large workloads (≥10K patterns): Maximum prefilter benefits  
export JAVA_OPTS="-Drmatch.engine=fastpath -Drmatch.prefilter.threshold=5000 -XX:+TieredCompilation -XX:CompileThreshold=500"

Documentation

See FASTPATH_PERFORMANCE_ANALYSIS.md for:

• Complete performance analysis and validation methodology
• JIT technique explanations and benchmark results
• Component-by-component optimization breakdown

---

Dispatch Optimization Experiments

rmatch includes benchmarks to test modern Java language features for dispatch pattern optimization on Java 25.

Quick Start

Run dispatch optimization benchmarks:

make bench-dispatch

Or run the script directly:

scripts/run_dispatch_benchmarks.sh

What's Tested

The benchmarks evaluate three optimization strategies:

1. Pattern Matching for instanceof - Java 16+ pattern matching vs traditional cast
2. Switch Expressions - Enhanced switch with arrow syntax vs if-else chains
3. Enum Dispatch - Switch expressions vs if-else for enum handling

Key Findings

Based on empirical testing:

• ✅ Enum switch expressions: 13.6% faster than if-else chains - RECOMMENDED
• ❌ Pattern matching instanceof: No measurable benefit (0.08%)
• ✅ If-else for char ranges: 19.5% faster than switch - keep current approach

Documentation

See DISPATCH_OPTIMIZATION_RESULTS.md for:

• Detailed benchmark results
• Performance analysis
• Specific recommendations for code changes
• Examples of patterns to refactor

These experiments follow the same methodology as GC experiments to provide data-driven guidance on whether modern language features improve performance.