Orca v2.0! Classic Meets Even More Modern: A (Slightly) Pragmatic Learning-Based Congestion Control for the Internet
This project is based on the findings from the SIGCOMM paper "Classic Meets Modern: A Pragmatic Learning-Based Congestion Control for the Internet" (https://dl.acm.org/doi/pdf/10.1145/3387514.3405892). This repository is forked from https://github.com/Soheil-ab/Orca, which provided scaffolding for our modifications.
This project extends the original Orca implementation with the following key improvements:
- Port to Modern Linux Kernel: We ported the original Orca kernel patches from Linux 4.13 to Linux Kernel 6.16, enabling deployment on modern systems.
- Dynamic Algorithm Switching: We implemented an in-kernel meta-controller that dynamically switches the underlying base congestion control algorithm (e.g., between TCP Cubic, Westwood, and BBR) based on real-time classification of the network environment (Wireless, Datacenter, High-BDP, etc.).
- Dynamic Monitoring Time Period (MTP): We replaced the fixed 20ms control loop with a dynamic MTP that adapts to the flow's Round-Trip Time (RTT). This synchronizes the agent's decision frequency with the network's feedback loop, improving stability on high-latency paths.
-
Modified Reward Function: We replaced the original linear reward function with a logarithmic utility function (
$R = \alpha \log(\text{Throughput}) - \beta \text{Delay} - \gamma \text{Loss}$ ) to improve stability and fairness.
- Usage of eBPF probes over the creation of the new networking module (DeepCC).
- Training and finetuning an LSTM model over the current MLP.
- Creating a three-stage model pipeline for MTP prediction, fallback algorithm prediction, and cwnd prediction.
The final project report and presentation slides can be found in /documentation.
The Orca 1.0 README (https://github.com/Soheil-ab/Orca) is copied below.
This release presents the source code and materials used for the experiments in our SIGCOMM'20 paper: "Classic Meets Modern: A Pragmatic Learning-Based Congestion Control for the Internet" (https://dl.acm.org/doi/abs/10.1145/3387514.3405892).
For the source code of our DeepCC plug-in reported in our IEEE JSAC'21 paper: "Wanna Make Your TCP Scheme Great for Cellular Networks? Let Machines Do It for You!", please checkout https://github.com/Soheil-ab/DeepCC.v1.0/ (The paper can be accessed here: https://ieeexplore.ieee.org/document/9252929)
Install Mahimahi (http://mahimahi.mit.edu/#getting)
```sh
sudo apt-get install build-essential git debhelper autotools-dev dh-autoreconf iptables protobuf-compiler libprotobuf-dev pkg-config libssl-dev dnsmasq-base ssl-cert libxcb-present-dev libcairo2-dev libpango1.0-dev iproute2 apache2-dev apache2-bin iptables dnsmasq-base gnuplot iproute2 apache2-api-20120211 libwww-perl
git clone https://github.com/ravinet/mahimahi
cd mahimahi
./autogen.sh && ./configure && make
sudo make install
sudo sysctl -w net.ipv4.ip_forward=1
```
- Python >= 3.4
- Tensorflow 1.14.0
mkdir ~/venv
sudo apt update
sudo apt install python3-pip
sudo pip3 install -U virtualenv
virtualenv ~/venv -p python3
If you change the path of the virtual environment, you need to change the source code accordingly. The source code assumes that virtual environment is set @ ~/venv .
source ~/venv/bin/activate
pip install --upgrade pip
pip install gym
pip install tensorflow==1.14
pip install sysv_ipc
Verify Installation
python -c "import tensorflow as tf; tf.enable_eager_execution(); print(tf.reduce_sum(tf.random_normal([1000, 1000])))"
To deactivate venv
(venv) $ deactivate
Simplest option to install Orca's patched Kernels is to install the prepared debian packages:
cd linux
sudo dpkg -i linux-image*
sudo dpkg -i linux-header*
sudo reboot
If you have already done the option 1, skip this part! Another option is to compile your own kernel using the provided patch. You can use the instructions provided here to do that: https://github.com/Soheil-ab/C2TCP-IFIP/
The source code is available in linux folder (https://github.com/Soheil-ab/Orca/blob/master/linux/linux-4-13-1-orca-0521%2Bc2tcp.patch)
After installing the Orca's kernel and restarting your system, use the following command to make sure that system is using the new kernel:
uname -r
The output should be 4.13.1-0521*. If not, you need to bring the 4.13.1-0521* Kernel image on top of the grub list. For instance, you can use grub-customizer application. Install the grub-customizer using following:
sudo add-apt-repository ppa:danielrichter2007/grub-customizer
sudo apt-get update
sudo apt-get install grub-customizer
sudo grub-customizer
To build the required applications, run the following:
./build.sh
This release includes two versions: A standalone actor version and an actor-learner version. In the Standalone actor version, no learner will be initiated. This can be usefull when you simply wanna test the current model over an emulated link. However, the actor-learner version requires a learner being initiated before any actor can be started. The second mode is usefull for learning a new/better model and also, it still can be used for performing a simple test over emulated links.
./orca-standalone-emulation.sh 44444
./orca.sh 4 44444
Results will be generated automatically in rl-module/log/* You can check out the summary of results at rl-module/log/sum-*
- To distribute the actors over remote servers, you need to change "learner_ip" and "actor_ip" fields of the params.json file to the corresponding server IPs.
- Start the learner first.
- Set
remote:truewhen using remote servers. - Set
num_actors: N, N is the number of actors.
To use traces avaialbe @ https://github.com/Soheil-ab/Cellular-Traces-NYC , copy them to the traces folder of the project.