Skip to content

prefetching.md

Latest commit

 

History

History
191 lines (144 loc) · 4.33 KB

prefetching.md

File metadata and controls

191 lines (144 loc) · 4.33 KB
Raw

Prefetching

The Cache Simulator supports multiple prefetching strategies to reduce cache miss latency.

Overview

Prefetching brings data into the cache before it is explicitly requested, hiding memory latency for predictable access patterns.

Prefetching Methods

Stream Buffer Prefetching

Sequential access detection and prefetching.

How it works:

  1. Detect sequential access patterns
  2. Prefetch subsequent cache lines
  3. Store in dedicated stream buffers

Configuration:

{
  "l1": {
    "prefetching": {
      "enabled": true,
      "type": "stream",
      "distance": 4,
      "bufferSize": 4
    }
  }
}

Parameters:

Parameter Description Default
distance Number of blocks to prefetch ahead 4
bufferSize Stream buffer entries 4

Stride Prefetching

Pattern-based prefetching for regular stride accesses.

How it works:

  1. Track address deltas between accesses
  2. Detect repeating stride patterns
  3. Predict and prefetch next addresses

Configuration:

{
  "l1": {
    "prefetching": {
      "enabled": true,
      "type": "stride",
      "confidence": 2,
      "tableSize": 256
    }
  }
}

Parameters:

Parameter Description Default
confidence Required pattern matches before prefetch 2
tableSize Stride predictor table entries 256

Adaptive Prefetching

Dynamic strategy selection based on workload.

How it works:

  1. Monitor prefetch accuracy
  2. Adjust aggressiveness based on feedback
  3. Switch strategies when patterns change

Configuration:

{
  "l1": {
    "prefetching": {
      "enabled": true,
      "adaptive": true,
      "minAccuracy": 0.5,
      "aggressiveness": "medium"
    }
  }
}

Aggressiveness levels:

  • conservative: Prefetch only on high confidence
  • medium: Balance between coverage and accuracy
  • aggressive: Prefetch on any pattern detection

Command Line Usage

# Enable prefetching (default settings)
./cachesim 64 32768 4 262144 8 1 4 traces/sequential.txt
#                              P  D = Prefetch enabled, Distance 4

# Disable prefetching
./cachesim 64 32768 4 262144 8 0 0 traces/random.txt

Prefetch Statistics

The simulator reports prefetching metrics:

Stream Buffer Statistics:
  Buffer Size: 4
  Hits: 1234
  Accesses: 5000
  Hit Rate: 24.7%
  
Prefetch Statistics:
  Prefetches Issued: 2500
  Useful Prefetches: 1850
  Accuracy: 74.0%
  Coverage: 46.3%

Key metrics:

Metric Description
Hits Accesses served from prefetched data
Accuracy Useful prefetches / Total prefetches
Coverage Prefetch hits / Total misses avoided

When to Use Prefetching

Enable prefetching for:

  • Sequential access patterns (array traversal)
  • Regular stride patterns (matrix operations)
  • Streaming workloads

Disable prefetching for:

  • Random access patterns
  • Small working sets
  • Memory-bandwidth constrained systems

Performance Impact

Typical improvements with prefetching:

Access Pattern Miss Reduction Notes
Sequential 50-80% Best case for stream buffers
Strided 30-60% Depends on stride regularity
Random 0-5% Prefetching not effective
Mixed 15-40% Adaptive works best

Best Practices

  1. Analyze workload first - Check access patterns before enabling
  2. Start conservative - Begin with small prefetch distance
  3. Monitor accuracy - Disable if accuracy drops below 50%
  4. Consider bandwidth - Aggressive prefetching uses more bandwidth
  5. Use adaptive mode - For workloads with varying patterns

Troubleshooting

Low prefetch accuracy:

  • Reduce prefetch distance
  • Switch to stride-based for irregular patterns
  • Enable adaptive mode

No improvement from prefetching:

  • Verify workload has predictable patterns
  • Check if working set fits in cache already
  • Increase cache size before tuning prefetch

Cache pollution from prefetching:

  • Reduce aggressiveness
  • Use dedicated prefetch buffer
  • Consider no-prefetch baseline

See Also