wifi-densepose/examples/scipix/docs/BENCHMARKS.md

ruvector-scipix Benchmark Suite

Comprehensive performance benchmarking for the Scipix OCR clone using Criterion.

Overview

This benchmark suite provides detailed performance analysis across all critical components of the OCR system:

• OCR Latency: End-to-end OCR performance metrics
• Preprocessing: Image preprocessing pipeline performance
• LaTeX Generation: LaTeX AST generation and string building
• Inference: Model inference benchmarks (detection, recognition, math)
• Cache: Embedding cache and similarity search performance
• API: REST API request/response handling
• Memory: Memory usage, growth, and fragmentation analysis

Performance Targets

Primary Targets

• Single Image OCR: < 100ms at P95
• Batch Processing (16 images): < 500ms total
• Preprocessing Pipeline: < 20ms
• LaTeX Generation: < 5ms

Secondary Targets

• Cache Hit Latency: < 1ms
• Similarity Search (1000 embeddings): < 10ms
• API Request Parsing: < 0.5ms
• Model Warm-up: < 200ms

Running Benchmarks

Run All Benchmarks

cd examples/scipix
./scripts/run_benchmarks.sh all

Run Specific Benchmark Suite

# OCR latency benchmarks
./scripts/run_benchmarks.sh latency

# Preprocessing benchmarks
./scripts/run_benchmarks.sh preprocessing

# LaTeX generation benchmarks
./scripts/run_benchmarks.sh latex

# Model inference benchmarks
./scripts/run_benchmarks.sh inference

# Cache benchmarks
./scripts/run_benchmarks.sh cache

# API benchmarks
./scripts/run_benchmarks.sh api

# Memory benchmarks
./scripts/run_benchmarks.sh memory

Quick Benchmark Suite

For rapid iteration during development:

./scripts/run_benchmarks.sh quick

CI Benchmark Suite

Minimal samples for continuous integration:

./scripts/run_benchmarks.sh ci

Baseline Tracking

Save Current Results as Baseline

BASELINE=v1.0 ./scripts/run_benchmarks.sh all

Compare with Saved Baseline

./scripts/run_benchmarks.sh compare v1.0

Compare with Main Branch

BASELINE=main ./scripts/run_benchmarks.sh all
./scripts/run_benchmarks.sh compare main

Benchmark Details

1. OCR Latency Benchmarks (ocr_latency.rs)

Tests end-to-end OCR performance across various scenarios:

• Single Image OCR: Different image sizes (224x224 to 1024x1024)
• Batch Processing: Batch sizes from 1 to 32 images
• Cold vs Warm Start: Model initialization overhead
• Latency Percentiles: P50, P95, P99 measurements
• Throughput: Images per second

Key Metrics:

• Mean latency
• P95/P99 latency
• Throughput (images/sec)
• Batch efficiency

2. Preprocessing Benchmarks (preprocessing.rs)

Image preprocessing pipeline performance:

• Individual Transforms: Grayscale, blur, threshold, edge detection
• Full Pipeline: Sequential preprocessing chain
• Parallel vs Sequential: Batch processing comparison
• Resize Operations: Nearest neighbor and bilinear interpolation

Key Metrics:

• Transform latency
• Pipeline total time
• Parallel speedup
• Memory overhead

3. LaTeX Generation Benchmarks (latex_generation.rs)

LaTeX code generation from AST:

• Simple Expressions: Fractions, powers, sums
• Complex Expressions: Matrices, integrals, summations
• AST Traversal: Tree depth impact on performance
• String Building: Optimization strategies
• Batch Generation: Multiple expressions

Key Metrics:

• Generation latency
• AST traversal time
• String concatenation efficiency

4. Inference Benchmarks (inference.rs)

Neural network model inference:

• Text Detection Model: Bounding box detection
• Text Recognition Model: OCR text extraction
• Math Model: Mathematical notation recognition
• Tensor Preprocessing: Image to tensor conversion
• Output Postprocessing: NMS, confidence filtering, CTC decoding
• Batch Inference: Multi-image processing
• Model Warm-up: Initialization overhead

Key Metrics:

• Inference latency per model
• Batch throughput
• Preprocessing overhead
• Postprocessing time

5. Cache Benchmarks (cache.rs)

Embedding cache and similarity search:

• Embedding Generation: Image to vector embedding
• Similarity Search: Linear and approximate nearest neighbor
• Cache Hit/Miss Latency: Lookup performance
• Cache Insertion: Add new entries
• Batch Operations: Multi-query performance
• Cache Statistics: Memory and efficiency metrics

Key Metrics:

• Embedding generation time
• Search latency (linear vs ANN)
• Hit/miss ratio impact
• Memory per embedding

6. API Benchmarks (api.rs)

REST API performance:

• Request Parsing: JSON deserialization
• Response Serialization: JSON encoding
• Concurrent Requests: Multi-client handling
• Middleware Overhead: Auth, logging, validation, rate limiting
• Error Handling: Error response generation
• End-to-End Request: Full request cycle

Key Metrics:

• Parse/serialize latency
• Middleware overhead
• Concurrent throughput
• Error handling time

7. Memory Benchmarks (memory.rs)

Memory usage and management:

• Peak Memory: Maximum usage during inference
• Memory per Image: Batch processing memory scaling
• Model Loading: Memory required for model initialization
• Memory Growth: Leak detection over time
• Fragmentation: Allocation/deallocation patterns
• Cache Memory: Embedding storage overhead
• Memory Pools: Pool vs heap allocation
• Tensor Layouts: HWC vs CHW memory impact

Key Metrics:

• Peak memory usage
• Memory growth rate
• Fragmentation level
• Pool efficiency

HTML Reports

Criterion automatically generates detailed HTML reports with:

• Performance graphs
• Statistical analysis
• Regression detection
• Historical comparisons

View Reports

After running benchmarks, open:

open target/criterion/report/index.html

Or for a specific benchmark:

open target/criterion/ocr_latency/report/index.html

Interpreting Results

Latency Metrics

• Mean: Average latency across all samples
• Median (P50): 50th percentile - half of requests are faster
• P95: 95th percentile - 95% of requests are faster
• P99: 99th percentile - 99% of requests are faster
• Standard Deviation: Variance in latency

Throughput Metrics

• Images/Second: Processing rate
• Batch Efficiency: Speedup from batching
• Sustainable Throughput: Max rate with <95% success

Regression Detection

Criterion detects performance regressions automatically:

• Green: Performance improved
• Yellow: Minor change (within noise)
• Red: Performance regressed

Memory Metrics

• Peak Usage: Maximum memory at any point
• Growth Rate: Memory increase over time
• Fragmentation: Memory layout efficiency

Best Practices

Running Benchmarks

1. Consistent Environment: Run on the same hardware
2. Quiet System: Close other applications
3. Multiple Samples: Use sufficient sample size (50-100)
4. Warm-up: Allow for JIT compilation and caching
5. Baseline Tracking: Save results for comparison

Analyzing Results

1. Focus on Percentiles: P95/P99 more important than mean
2. Check Variance: High variance indicates instability
3. Profile Outliers: Investigate extreme values
4. Memory Leaks: Monitor growth rate
5. Regression Limits: Set acceptable thresholds

Optimization Workflow

1. Baseline: Establish current performance
2. Profile: Identify bottlenecks
3. Optimize: Implement improvements
4. Benchmark: Measure impact
5. Compare: Verify improvement vs baseline
6. Iterate: Repeat until targets met

Continuous Integration

CI Benchmark Configuration

# .github/workflows/benchmark.yml
name: Benchmarks

on:
  pull_request:
  push:
    branches: [main]

jobs:
  benchmark:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v3
      - uses: actions-rs/toolchain@v1
        with:
          toolchain: stable

      - name: Run benchmarks
        run: |
          cd examples/scipix
          ./scripts/run_benchmarks.sh ci

      - name: Compare with baseline
        run: |
          cd examples/scipix
          ./scripts/run_benchmarks.sh compare main

Troubleshooting

Benchmarks Running Slowly

• Reduce sample size: cargo bench -- --sample-size 10
• Use quick mode: ./scripts/run_benchmarks.sh quick
• Run specific benchmarks only

Inconsistent Results

• Ensure system is idle
• Disable CPU frequency scaling
• Run with higher sample size
• Check for thermal throttling

Memory Issues

• Monitor system memory during benchmarks
• Use memory profiling tools (valgrind, heaptrack)
• Check for memory leaks with growth benchmarks

Contributing

When adding new features:

1. Add corresponding benchmarks
2. Set performance targets
3. Run baseline before/after changes
4. Document any performance impact
5. Update this documentation

Resources

• Criterion.rs Documentation
• Rust Performance Book
• Benchmarking Best Practices