DeLFS: A Decentralized Log-Structured File System for Manycores

This repository contains the DeLFS source code, comparison file systems, formatting tools, and evaluation scripts.

Contents

• Repository structure
• Build and install DeLFS
• Run the benchmarks

Repository structure

DeLFS/
├── DeLFS/
│   ├── linux-6.6.8/
│   └── module_DeLFS/
├── DeLFS-CrashMonkey/
├── DeLFS-Git-Aging-Benchmark/
├── delfs-mkfs/
├── F2FSJ/
│   ├── f3fsj-6.6.8/
│   ├── f3fsj-6.6.8-SC/
│   └── f3fsj-6.6.8-SC-SF/
├── MAX/
├── ScaleCache/
│   ├── f2fs-6.6.8-SC/
│   └── f2fs-6.6.8-SC-SF/
├── ScaleLFS/
│   ├── ScaleLFS-6.6.8/
│   ├── ScaleLFS-6.6.8-SC/
│   └── ScaleLFS-6.6.8-SC-SF/
└── script/
    ├── filebench/
    ├── fio/
    └── RocksDB/

Each configuration under F2FSJ/ contains linux-6.6.8/ and module_f3fsj/. Each configuration under ScaleCache/ and ScaleLFS/ contains linux-6.6.8/ and module_f3fs/.

The commands below start from the repository root unless stated otherwise.

Build and install DeLFS

1. Build and install the kernel

cd DeLFS/linux-6.6.8
cp config-scalecache .config
make oldconfig

make -j "$(nproc)"
make -j "$(nproc)" modules
sudo make -j "$(nproc)" modules_install INSTALL_MOD_STRIP=1
sudo make -j "$(nproc)" install

cd ../..

Reboot into the installed kernel before building the DeLFS module. Check the running kernel with:

uname -r

2. Build the formatting tool

cd delfs-mkfs
./autogen.sh
./configure
make -j "$(nproc)"
cd ..

3. Format the SSD

DeLFS uses an Intel SSD in this setup. Identify the intended device before formatting it:

cd DeLFS
./check_model

Pass the device name without /dev/ to mkfs.sh. For example, if check_model identifies the intended Intel SSD as /dev/nvme0n1:

./mkfs.sh nvme0n1

Formatting destroys the existing data on the selected device. Verify the device name before running mkfs.sh.

4. Build and load the DeLFS module

From DeLFS/:

cd module_DeLFS
make -j "$(nproc)"

sudo modprobe -v lz4_compress
sudo modprobe -v lz4hc_compress
sudo modprobe -v jbd2
sudo insmod delfs.ko

5. Mount DeLFS

The supplied Filebench and RocksDB workloads use /mnt/test. Mount the formatted SSD there, replacing nvme0n1 with the device selected above:

sudo mkdir -p /mnt/test
sudo mount -t delfs /dev/nvme0n1 /mnt/test
findmnt --mountpoint /mnt/test

cd ../..

At this point, the working directory is the repository root.

Run the benchmarks

fio: microbenchmark

cd script/fio
sudo fio rand-write.fio

Filebench: macrobenchmarks

The repository provides three Filebench profiles used for the fileserver, varmail, and videoserver workloads. All three profiles use /mnt/test and specify a 60-second run.

Install Filebench separately if it is not already available. On the evaluation machine, its executable is /usr/local/bin/filebench.

From the repository root:

cd script/filebench
findmnt --mountpoint /mnt/test

sudo /usr/local/bin/filebench -f myfileserver.f
sudo /usr/local/bin/filebench -f myvarmail.f
sudo /usr/local/bin/filebench -f myvideoserver.f

Run each profile as a separate experiment and prepare the file system in the same initial state when comparing results.

RocksDB with YCSB: application benchmark

The RocksDB scripts store data under /mnt/test/ycsb-rocksdb-data. Mount DeLFS at /mnt/test before running them. Use sudo because the mounted directory requires root permission in this setup.

From the repository root:

cd script/RocksDB
findmnt --mountpoint /mnt/test

sudo ./load_rocksdb.sh
sudo ./run_rocksdb.sh

Run the load phase before the run phase. The paper evaluates YCSB workloads A, B, F, and update-only with 1–128 client threads, 10 million records, 10 fields of 4 KB per record, and 1.28 million operations per workload. Check the RocksDB scripts and their workload configuration files to select and reproduce each of these cases.