This is a testing repository of dora nodes that implements several cpu sync and async threadpool. Feel free to use it, to figure out what works best for you. This repository might not be up to date.
# Download the data
mkdir data
mkdir bin
# Clone dora in parent folder if it is not almounted there
cd ../
git clone https://github.com/dora-rs/dora.git
# Get an executable coordinator from `dora`
cd ./dora
cargo build --manifest-path coordinator/Cargo.toml --release
cd ../dora-sync-async-examples
cp ../dora/target/release/dora-coordinator ./bin/dora-coordinator
# Setup a Jaeger suite to collect metrics and traces
docker run -d -p6831:6831/udp -p6832:6832/udp -p16686:16686 jaegertracing/all-in-one:latest- To run the dataflow you can then run:
# The running command
./bin/dora-coordinator run ./dataflow.yml
If you go into the common crate, you will be able to change the run functions you want to benchmark. Those functions will be tested on all threadpool nodes defined in the dataflow.
To fetch the metric from Jaeger, I wrote a histogram.py that will collect the latest data in a tabular form from Jaeger API. The script looks something like this:
import pandas as pd
link = "http://localhost:16686/api/traces?service=sender&lookback=6h&prettyPrint=true&limit=1"
df = pd.read_json(link)
df = pd.DataFrame(df.data[0]["spans"])You will have to install the following dependency to run the python script
pip install jupyterthemes
python histogram.py
You can also check the Jaeger UI: http://localhost:16686
Fixing onnxruntime-rs shared library issues requires that you link the downloaded onnxruntime shared library in your target build folder:
export ORT_USE_CUDA=1
export LD_LIBRARY_PATH="PATH TO DORA_CPU_BOUND..."/cpu-bound-async/target/release/build/onnxruntime-sys-186188f4edb1a21e/out/onnxruntime/onnxruntime-linux-x64-gpu-1.8.0/lib:${LD_LIBRARY_PATH}
Feel free to go on haixuantao/onnxruntime-rs on branch owned_environment if you have any issues with the GPU.