This repository contains the prototype for our ICNP 2026 paper, GRC: Extending RDMA to Wide-Area Networks via Gateway-Mediated Rate Control.
GRC enables RDMA over shared wide-area networks by moving WAN rate control to the gateway connecting a data center to the WAN.
- Gateway-mediated control: estimates WAN delay and maintains a reference rate for each destination data center.
- Fast RNIC reaction: translates gateway decisions into standard CNPs that existing RoCEv2 RNICs can process.
- Minimal infrastructure changes: requires gateway changes without RDMA-specific support in WAN devices.
- Compatible with intra-DC transports: WAN control is separated from the congestion-control mechanism used inside each data center.
NS3/: complete NS-3 simulator, GRC implementation, topologies, workload generators, experiment runners, and analysis utilities.P4/: P4 data-plane and C control-plane source for the switch prototype.
The simulator was tested on Ubuntu 20.04.
sudo apt install build-essential python3 libgtk-3-0 bzip2 python2
python3 -m pip install numpy matplotlib pandas cycler
cd NS3
./waf configure --build-profile=optimized
./wafThe remaining simulation commands are run from the NS3/ directory.
Run a small GRC experiment:
python3 run.py \
--simul_time 0.05 \
--cdf WebSearch \
--intra_load 30 \
--inter_load_all 60 \
--my_flow '' \
--tcp_flow '' \
--wan_cc_mode 1 \
--msg smoke-grcEach run stores its configuration and output under mix/output/. Use
python3 check.py state to inspect running and recently completed experiments.
Generate the WebSearch inputs for the main load sweep:
for d in 0 60 120 180; do
python3 config/large_traffic_gen.py -f w -b 150 -d "$d"
doneThe overall runner is kept in the simulator tree. Inspect or launch the sweep
from NS3/:
cd NS3
python3 expr-overall/autorun.py --dry_run --sleep 0
python3 expr-overall/autorun.py --sleep 1The generator is randomized, so regenerated inputs may differ between runs. Experiment output is not included in this repository because of its size.
For mixed RDMA/TCP traffic, pass the included TCP input to run.py:
python3 run.py \
--topo cernet_topo \
--simul_time 0.1 \
--my_flow w-dynamic-150-180 \
--tcp_flow config/w-tcp-100.txt \
--wan_cc_mode 1Use analysis/deep_analyse.py or
analysis/experiment_report_template.ipynb to inspect completed runs.
P4/dataplane/ and P4/controlplane/ contain the prototype source. Tofino
SDK files, SDE binaries, and hardware experiment output are not included.
Jue Zhang, Menghao Zhang, Zihan Niu, Bo Peng, Shucan Yang, and Xiaohe Hu. GRC: Extending RDMA to Wide-Area Networks via Gateway-Mediated Rate Control. IEEE International Conference on Network Protocols (ICNP), 2026.
@inproceedings{zhang2026grc,
author = {Jue Zhang and Menghao Zhang and Zihan Niu and Bo Peng and
Shucan Yang and Xiaohe Hu},
title = {GRC: Extending RDMA to Wide-Area Networks via Gateway-Mediated Rate Control},
booktitle = {IEEE International Conference on Network Protocols (ICNP)},
year = {2026}
}