This repository contains source code and scripts for reproducing key results from the cache_ext paper (SOSP 2025) for the purposes of artifact evaluation.

If using cache_ext, please cite the reference below.

We use a Cloudlab instance of type c6525-25g running Ubuntu 22.04, with a maxed out temporary disk.

There are four major components:

• A modified Linux kernel based on Linux v6.6.8 that supports cache_ext. This also includes supporting changes to libbpf and bpftool.
• Policies: A set of custom page cache eviction policies implemented in eBPF.
• LevelDB with minor changes to support cache_ext experiments.
• My-YCSB: an efficient C++ YCSB benchmarking framework.

Additionally, there are several other experiments that make use of pre-existing tools with no modifications, such as fio and ripgrep.

cache_ext
|-- policies/                   : eBPF page cache eviction policies
|-- bench/                      : Benchmarking framework
|-- eval/                       : Evaluation scripts and result analysis
|   |-- cpu-overhead/           : CPU overhead experiment
|   |-- filesearch/             : File search experiment
|   |-- get-scan/               : GET-SCAN experiment
|   |-- isolation/              : Isolation (per-cgroup) experiment
|   |-- twitter/                : Twitter trace experiment
|   \-- ycsb/                   : YCSB experiment
|-- utils/                      : Helper scripts
\-- *.sh                        : Component installation and build scripts

First, clone the repo into Cloudlab's temporary disk (i.e., /mydata) and initialize the submodules:

cd /mydata
git clone https://github.com/cache-ext/cache_ext.git
cd cache_ext
git submodule update --init --recursive

Next, you must compile and install the custom Linux kernel:

./install_kernel.sh

This will also set up libbpf and bpftool.

After the kernel is compiled and installed, you will be prompted to reboot into cache_ext kernel:

sudo grub-reboot "Advanced options for Ubuntu>Ubuntu, with Linux 6.6.8-cache-ext+"
sudo reboot now

Some of the remaining components can only be compiled on the cache_ext kernel, so you must wait for the system to reboot and then log back in.

Then, run the following command to download the various datasets used. The datasets are stored on a public Google Cloud bucket. These will be set up on Cloudlab. The download step may take a while, as the datasets are hundreds of gigabytes in size.

cd /mydata/cache_ext
./download_dbs.sh

Then, you can build and install the other components:

./install_filesearch.sh
./install_leveldb.sh
./install_misc.sh
./install_ycsb.sh
./setup_isolation.sh
./build_policies.sh

Note that this will require approximately 500GB of disk space.

You can run the experiments in the eval directory. Each subdirectory contains a run.sh script that sets up the environment and runs the experiment. For example, to run the file search benchmark, you can execute:

cd /mydata/cache_ext/eval/filesearch
./run.sh

Note that by default, the run.sh scripts are set to run each experiment three times to get an average result. You can change the number of iterations by modifying the ITERATIONS variable in each run.sh script.

The results will be saved in a top-level results directory as JSON files.

Many of the experiments will take several hours to run due to the number of configurations, policies, and iterations. For example, by default, the YCSB benchmark will take approximately 20 hours to run. We recommend using the screen command to start a persistent session that can run the command in the background and avoid losing progress if the SSH connection is interrupted.

To exit a screen session, you can press Ctrl + A followed by Ctrl + D.

We include a Jupyter notebook bench/bench_plot.ipynb that can be used to plot the results of the experiments. You can run this notebook in a Jupyter environment (i.e., Jupyter Notebook or VSCode).

For example, to start up a Jupyter server, you can run:

cd /mydata/cache_ext/bench
jupyter notebook

This will start a Jupyter server and open a web interface in your browser. You can then open the bench_plot.ipynb notebook and run the cells to generate the plots.

The figures will be saved as PDFs in a created figures/ directory.

If using cache_ext, please include the following citation:

@inproceedings{cacheext,
author = {Zussman, Tal and Zarkadas, Ioannis and Carin, Jeremy and Cheng, Andrew and Franke, Hubertus and Pfefferle, Jonas and Cidon, Asaf},
title = {cache_ext: Customizing the Page Cache with eBPF},
year = {2025},
isbn = {9798400718700},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3731569.3764820},
doi = {10.1145/3731569.3764820},
pages = {462–478},
numpages = {17},
keywords = {operating systems, eBPF, page cache},
location = {Lotte Hotel World, Seoul, Republic of Korea},
series = {SOSP '25}
}