ATP-BytePS is the endhost training system developed for the paper in NSDI21'. This is based on BytePS v0.2.4 substituting PS-lite communication library to ATP communication library.
git clone -b p4ml_024 git@github.com:laochanlam/byteps.git
cd byteps
sudo -E BYTEPS_WITHOUT_MXNET=1 BYTEPS_WITHOUT_TENSORFLOW=1 python setup.py install
The dependencies can be referred to BytePS v0.2.4.
sudo -E CUDA_VISIBLE_DEVICES=0 DMLC_WORKER_ID=0 DMLC_NUM_WORKER=2 DMLC_INTERFACE=enp178s0f0 DMLC_ROLE=worker DMLC_NUM_SERVER=1 DMLC_PS_ROOT_URI=192.168.0.3 DMLC_PS_ROOT_PORT=6767 EVAL_TYPE=benchmark P4ML_APP=1 python byteps/launcher/launch.py python byteps/example/pytorch/benchmark_byteps.py --model vgg16 --num-iters 10
sudo -E CUDA_VISIBLE_DEVICES=0 DMLC_WORKER_ID=1 DMLC_NUM_WORKER=2 DMLC_INTERFACE=enp178s0f0 DMLC_ROLE=worker DMLC_NUM_SERVER=1 DMLC_PS_ROOT_URI=192.168.0.3 DMLC_PS_ROOT_PORT=6767 EVAL_TYPE=benchmark P4ML_APP=1 python byteps/launcher/launch.py python byteps/example/pytorch/benchmark_byteps.py --model vgg16 --num-iters 10
$ cd $ATP_REPO/server/
$ make
# Usage: ./app [AppID]
sudo ./app 1
If you are using physical switch, compile the switch program then jump to Terminal 2 directly.
$ cd $SDE
$ $TOOLS/p4_build.sh ~/git/p4ml/p4src/p4ml.p4
# (Optional) for software Tofino behavior model
$ ./run_tofino_model.sh -p p4ml
$ cd $SDE
$ ./run_switchd.sh -p p4ml
$ $SDE/run_p4_tests.sh -t $ATP_REPO/ptf/ -p p4ml
$ $TOOLS/run_pd_rpc.py -p p4ml $ATP_REPO/run_pd_rpc/setup.py
- [NSDI'21] "ATP: In-network Aggregation for Multi-tenant Learning". ChonLam Lao, Yanfang Le, Kshiteej Mahajan, Yixi Chen, Wenfei Wu, Aditya Akella, Michael Swift.
Any questions? Please feel free to reach us at inatpcontact@gmail.com. You are more likely to receive a helpful response if your question is specific, self-contained and concise.
This repository is modified based on BytePS v0.2.4.