Reproducibility Study of CACHEUS on CloudVPS

This repository is an academic reproducibility study based on the original CACHEUS cache replacement project from FAST 2021. It keeps the original simulator and algorithms, then adds a focused CloudVPS-only pipeline for preparing traces, running experiments, and generating summary results.

Scope

This study evaluates:

• Algorithms: lru, lfu, arc, lirs, lecar, cacheus
• Cache sizes: 0.001, 0.005, 0.01, 0.05, 0.1 of unique requests
• Dataset: public CloudVPS traces only

This is not a full reproduction of every dataset in the original paper. It is a smaller academic reproduction focused on the CloudVPS workload.

Repository Contents

• code/run.py: simulator entry point.
• code/algs/: CACHEUS and baseline cache replacement algorithms.
• cloudvps_pipeline.py: offline CloudVPS reproduction pipeline.
• results/cloudvps/: curated summary CSVs and figures from the reproduction.

Raw traces, decoded traces, intermediate work files, and full logs are not tracked in git.

Requirements

• Python 3
• numpy
• matplotlib
• scipy
• blkparse from Linux blktrace tools

Install Python dependencies:

pip install -r requirements.txt

On Debian/Ubuntu:

sudo apt-get install blktrace

Run

Place CloudVPS vps*.tar.gz archives, or extracted vps* directories, under:

data/cloudvps_raw/

Then run:

python cloudvps_pipeline.py

Optional custom paths:

python cloudvps_pipeline.py --raw-dir data/cloudvps_raw --work-dir work/cloudvps --out-dir results/cloudvps

The pipeline extracts traces, decodes .blktrace.0 files with blkparse, converts them to .blk, runs the simulator, and writes CSV summaries and plots.

Results

Curated results are stored in results/cloudvps/, including:

• summary_by_algorithm.csv
• summary_by_cache_size.csv
• cacheus_vs_baselines.csv
• cacheus_delta_summary.csv
• cacheus_paired_t_tests.csv
• conversion_summary.csv
• run_metadata.txt
• figures/*.png for hit-rate, runtime, heatmap, CACHEUS-vs-baseline, and paired-test plots

In this CloudVPS-only reproduction, CACHEUS performs strongly against LRU, LFU, ARC, and LeCaR, and is nearly tied with LIRS overall.

Statistical Analysis

CACHEUS is compared with each baseline using paired t-tests over matched (trace file, cache size) runs. The paired design controls for workload and cache-size effects because both algorithms are evaluated on the same trace and cache-size setting. The tests use hit-rate differences in percentage points.

The current paired-test results support CACHEUS outperforming LRU, LFU, ARC, and LeCaR on average. They do not show a statistically significant difference between CACHEUS and LIRS in this CloudVPS-only run.

Limitations

This repository is a focused CloudVPS reproduction, not a full reproduction of every dataset and experiment from the original CACHEUS paper. Raw CloudVPS traces, decoded traces, converted .blk files, full simulator logs, and local work directories are intentionally not tracked in git.

References

• Rodriguez et al., Learning Cache Replacement with CACHEUS, FAST 2021.
• Vietri et al., Driving Cache Replacement with ML-based LeCaR, HotStorage 2018.