BPF-oF: Storage Function Pushdown Over the Network

This repository contains source code and instructions to reproduce key results in the BPF-oF paper (arxiv). A draft of the paper is added to this repository.

Dependencies

• Linux 5.12.0 with BPF-oF changes: https://github.com/xrp-project/linux/tree/bpf-of
• RocksDB with BPF-oF changes: https://github.com/xrp-project/rocksdb/tree/bpf-of
• My-YCSB (benchmarking suite) with RocksDB support: https://github.com/xrp-project/My-YCSB/tree/bpf-of

Requirements

You need a different type of machine depending on the benchmark you want to run:

Configuration	Host Machine	Target Machine
TCP SSD	c6525-100g	c6525-100g
RDMA SSD	r6525	r6525
TCP Optane	c6525-100g	flex0{1-4}

Step 1: Get machines from CloudLab

NOTE: If machines were given to you, skip this skep.

1. To get machines from CloudLab, you need to create an experiment. A CloudLab experiment uses a profile as a template. Since BPF-oF uses specific profiles, you need to load them first. Go to CloudLab, click on Experiments -> Create Experiment Profile and create profile nvmeof-bench and profile optane-plus-client, located in folder cloudlab-profiles.

2. Create an experiment depending on the benchmarks you want to run:

   • For TCP SSD, use the nvmeof-bench profile.
   • For RDMA SSD, use the nvmeof-bench profile and change the machine type to r6525.
   • For TCP Optane, use the optane-plus-client profile.

3. Wait until CloudLab provisions the machines for the experiment.

Step 2: Setup the machine

TODO: Add Ansible instructions.

Step 3: Run benchmarks

TODO: Adapt the scripts to use this repo instead of the phd repo.

Running the benchmarks is a fully automated process.

TRANSPORT="tcp" # tcp | rdma | local
CLOUDLAB_EXPERIMENT="nvmeof-bench" # nvmeof-bench | optane-plus-client

python3 bench_rocksdb_supervisor.py \
    --cloudlab-experiment $CLOUDLAB_EXPERIMENT \
    --transport $TRANSPORT \
    --debug

Step 4: Get results

For NVMe-oF:

REMOTE_USERNAME="..."
HOST_MACHINE="..."

scp -r $REMOTE_USERNAME@$HOST_MACHINE:/mydata/bpfof_results ~/

For local:

REMOTE_USERNAME="..."
TARGET_MACHINE="..."

scp -r $REMOTE_USERNAME@$TARGET_MACHINE:/mydata/bpfof_results ~/

Step 5: Plot the results

TODO: Adapt python notebook.