task.txt

You are responsible for running all experiments required for a research paper evaluating an RL-based cache eviction algorithm compared to classical algorithms. You must run the following experiments, collect the specified metrics, generate the tables, and prepare all plots. Use the current RL implementation as-is unless otherwise specified.

1. Baseline Performance Evaluation

Run these algorithms on identical workloads:

LRU

LFU (optional)

ARC (optional)

RL-Hybrid (current RL model)

For each, record:

Hit rate (%)

Total runtime (seconds)

Average per-request latency (microseconds)

Latency p50, p90, p99

Memory usage (MB)

Output:

A baseline comparison table covering all metrics.

2. Microbenchmark Breakdown of RL

Break down RL-Hybrid runtime into the following components:

Time to construct candidate states

Time for model inference (1 candidate and k candidates)

Time for choosing eviction

Time for updating recency/frequency metadata

Output:

A table comparing each operation vs LRU’s equivalent O(1) operations.

A bar chart showing time distribution per component.

3. Sensitivity Analysis
3.1 Cache Size Variation

Run all algorithms on caches of size:
10, 20, 30, 50, 100, 200

Record for each:

Hit rate

Runtime

RL overhead vs LRU (%)

3.2 Candidate Size Variation

For RL, run with:
k = 4, 8, 16, 32

Record:

Hit rate

Runtime

Scaling of inference time

3.3 Workload Distribution Variation

Test the following workloads:

Zipf distributions with α = 0.5, 0.8, 1.0, 1.2, 1.5

Uniform random

Gaussian phase workloads

Bursty workloads

Any real-world trace if available

For each workload, record:

Hit rate

Runtime

RL slowdown factor vs LRU

Output:

Sensitivity tables for cache size, k, and workloads.

4. Workload Behavior Tests
4.1 Long-Term Patterns

Use periodic and phase-shift workloads to test adaptation.
Record:

RL hit rate trend over time

Does RL adapt or collapse?

4.2 Cold-Start Behavior

Record:

Hit rate during first 5k accesses

RL stabilization time

4.3 Adversarial Workloads

Run workloads known to break classical heuristics and record RL vs LRU performance.

Output:

One table summarizing behavioral responses for each workload type.

5. RL Model Internal Behavior Analysis
5.1 Weight Inspection

Extract RL model weights.
Determine:

Relative significance of recency vs frequency vs rank.

Whether RL behaves like LRU or LFU.

5.2 Q-Value Analysis

Collect:

Histogram of Q-values

Q-value variance

Q-value collapse indicators

5.3 Training Dynamics

Record:

Loss curve over training

Hit rate during training

5.4 Feature Correlation

Compute correlation between:

RL predicted scores

True LRU recency ranks

Output:

Weight inspection table

Correlation table

6. Ablation Studies

Run the following RL variants:

6.1 Remove Frequency Feature

State = [recency, rank]

6.2 Remove Rank Feature

State = [recency, frequency]

6.3 Linear Model Instead of MLP

Replace MLP with single linear layer.

6.4 Smaller Networks

Test architectures:

32–32–1

16–16–1

8–8–1

6.5 Pure Weighted Heuristic

Use:
score = αrecency + βfrequency + γ*rank
(no RL)

Record for each:

Hit rate

Runtime

Parameter count

Output:

Ablation table comparing all variants.

7. Statistical Significance
Run every experiment 10 times

For each:

Compute mean

Standard deviation

95% confidence intervals

Perform t-tests comparing:

LRU vs RL hit rate

LRU vs RL latency

Output:

Statistical significance table

Boxplots or error-bar graphs

8. Required Plots

You must generate all the following:

Hit rate vs cache size (line graph)

Runtime vs cache size

Runtime vs k

Hit rate vs Zipf α

Q-value distribution histogram

Training loss curve

Feature weight bar chart

Latency CDF plot (LRU vs RL)

9. Final Summary Tables
9.1 Performance Summary Table

Columns:
Algorithm | Hit Rate | Runtime | p99 Latency | Memory Usage | Cache Size

9.2 Sensitivity Table

Cache Size | k | LRU Hit | RL Hit | RL Latency | Notes

9.3 Workload Behavior Table

Workload Type | LRU Hit | RL Hit | RL Slowdown | Behavior Summary

9.4 Ablation Table

Variant | Hit Rate | Runtime | Params | Notes

9.5 Model Explanation Table

Feature | Weight | Importance | Interpretation

10. Values to Extract for Paper

You must extract:

Hit rate

Runtime

Latency percentiles

Q-value variance

Model parameter count

Inference time (microseconds)

Correlation with LRU ranking

Warm-up time

Memory overhead

Scaling factors