跳到主要内容
版本：🚧 开发中

Big Batch Test Plan (High Concurrency C=50+)

Project: Issue #995 - ModernBERT-base-32k Integration
Required: NVIDIA A100 GPU (40GB+ VRAM)

Overview​

This test plan covers validation of ModernBERT-base-32k under high concurrency (big batch) scenarios. These tests cannot be completed with the current environment (NVIDIA L4 GPU, 23GB VRAM) due to memory fragmentation issues and require an A100 GPU with 40GB+ VRAM.

Infrastructure Status: Ready - All tools and test frameworks are prepared

Test Requirements​

Hardware Requirements​

  • GPU: NVIDIA A100 (40GB+ VRAM) - Required
  • System RAM: 64GB+ recommended
  • CUDA: Version 12.0+
  • Driver: Latest NVIDIA driver

Software Requirements​

  • benchmark_concurrent.rs - Supports C=50, C=100 (currently fails due to memory)
  • benchmark_performance.rs - Performance profiling tool
  • Flash Attention 2 enabled
  • All dependencies installed

Test Cases​

1. High Concurrency Testing (C=50, C=100)​

1.1 Low Context Length (1K-4K tokens)​

Context LengthConcurrencyExpected Success RateExpected Latency
1024 tokensC=50≥ 90%< 2000ms (p95)
1024 tokensC=100≥ 80%< 3000ms (p95)
4096 tokensC=50≥ 80%< 15000ms (p95)
4096 tokensC=100≥ 70%< 20000ms (p95)

Test Steps:

  1. Run benchmark_concurrent.rs with C=50, C=100
  2. Test with 1K, 4K tokens
  3. Measure latency (mean, p50, p95, p99)
  4. Track success/error rates
  5. Document memory usage

Deliverables:

  • Latency statistics for each concurrency level
  • Success/error rates
  • Memory usage profiles
  • Throughput measurements

1.2 Medium Context Length (8K tokens)​

Context LengthConcurrencyExpected Success RateExpected Latency
8192 tokensC=50≥ 70%< 25000ms (p95)
8192 tokensC=100≥ 60%< 35000ms (p95)

Test Steps:

  1. Run benchmark_concurrent.rs with C=50, C=100
  2. Test with 8K tokens
  3. Measure latency (mean, p50, p95, p99)
  4. Track success/error rates
  5. Document memory usage and fragmentation

Deliverables:

  • Latency statistics
  • Success/error rates
  • Memory fragmentation analysis
  • Recommendations for production

2. Throughput Analysis​

2.1 Requests Per Second (RPS)​

Test Steps:

  1. Measure throughput at different concurrency levels
  2. Calculate RPS for C=50, C=100
  3. Compare with C=1, C=10 baseline
  4. Document throughput scaling

Deliverables:

  • RPS measurements
  • Throughput scaling analysis
  • Bottleneck identification

3. Memory Fragmentation Analysis​

3.1 Memory Management Under High Concurrency​

Test Steps:

  1. Monitor GPU memory usage during high concurrency tests
  2. Track memory allocation/deallocation patterns
  3. Identify memory fragmentation issues
  4. Test memory cleanup strategies
  5. Document recommendations

Deliverables:

  • Memory usage profiles
  • Fragmentation analysis
  • Cleanup strategy recommendations

4. Latency Distribution Analysis​

4.1 P95/P99 Latency Under High Concurrency​

Test Steps:

  1. Measure latency distribution at C=50, C=100
  2. Analyze p50, p95, p99 percentiles
  3. Identify latency spikes
  4. Document tail latency behavior

Deliverables:

  • Latency distribution charts
  • P95/P99 analysis
  • Tail latency recommendations

5. Error Rate Analysis​

5.1 OOM Error Patterns​

Test Steps:

  1. Track OOM errors at different concurrency levels
  2. Analyze error patterns
  3. Identify failure modes
  4. Document error recovery strategies

Deliverables:

  • Error rate analysis
  • Failure mode documentation
  • Recovery strategy recommendations

Expected Outcomes​

Success Criteria​

  1. C=50:

    • 1K tokens: ≥ 90% success rate
    • 4K tokens: ≥ 80% success rate
    • 8K tokens: ≥ 70% success rate

  2. C=100:

    • 1K tokens: ≥ 80% success rate
    • 4K tokens: ≥ 70% success rate
    • 8K tokens: ≥ 60% success rate

  3. Latency:

    • P95 latency within acceptable limits
    • No excessive tail latency

  4. Memory:

    • Memory usage within GPU limits
    • No memory leaks
    • Acceptable fragmentation

Infrastructure Readiness​

Tools Ready​

  • benchmark_concurrent.rs - Supports C=50, C=100 (currently fails due to memory)
  • benchmark_performance.rs - Performance profiling ready
  • Flash Attention 2 enabled
  • All dependencies installed

How to Run Tests​

  1. Ensure sufficient GPU memory (A100 40GB+)

  2. Run benchmark with high concurrency:

    cargo run --example benchmark_concurrent --release --features cuda,flash-attn

  3. Monitor memory usage:

    watch -n 1 nvidia-smi

Resource Estimates​

Time Estimates​

  • High Concurrency Testing: 6-8 hours
  • Throughput Analysis: 2-3 hours
  • Memory Fragmentation Analysis: 3-4 hours
  • Latency Distribution Analysis: 2-3 hours
  • Error Rate Analysis: 2-3 hours

Total: ~15-21 hours

Resource Requirements​

  • GPU: A100 (40GB VRAM) - Required
  • System RAM: 64GB+ recommended
  • Storage: ~10GB for test data and results

Deliverables​

  1. Test Results Report

    • Latency measurements (C=50, C=100)
    • Success/error rates
    • Throughput measurements
    • Memory usage profiles

  2. Performance Analysis

    • Throughput scaling analysis
    • Memory fragmentation analysis
    • Latency distribution analysis
    • Error rate patterns

  3. Recommendations

    • Production concurrency limits
    • Memory management strategies
    • Scaling recommendations
    • Error handling strategies

Known Issues (From Current Testing)​

Memory Fragmentation​

  • Issue: GPU memory not released between tests
  • Impact: Subsequent high-concurrency tests fail
  • Workaround: Test with fresh GPU state or longer wait times
  • Solution: Investigate memory cleanup strategies on A100

OOM Errors at High Concurrency​

  • Issue: OOM errors at C=50+ for 8K+ tokens
  • Impact: Cannot test high concurrency with current environment
  • Solution: A100 with 40GB+ VRAM required

References​

  • Performance Validation: Performance Validation
  • Deployment Guide: Deployment Guide
  • Benchmark Tool: candle-binding/examples/benchmark_concurrent.rs
  • Performance Tool: candle-binding/examples/benchmark_performance.rs