Guiding Device Specific Optimization using Architecture-Independent Metrics

Installation

This project uses Docker to facilitate reproducibility. As such, it has the following dependencies:

• Cuda 9.0 Runtime -- available here
• Docker -- available here
• nvidia-docker2, install instructions found here
• Docker nvidia container, installed with: sudo apt install nvidia-container-runtime

Build

To generate a docker image named guiding-optimisation, run:

docker build -t guiding-optimisation .

Run

To start the docker image run:

docker run --runtime=nvidia -it --mount src=pwd,target=/tuning-hints-with-aiwc,type=bind -p 8888:8888 -p 9091:9091 --net=host adi/tuning-hints-with-aiwc

And run the codes with: cd /guiding-optimisation-with-aiwc/codes

make

make test

This generates a sample of the runtimes with libscibench and the AIWC metrics

For reproducibility, BeakerX has also been added for replicating results and for the transparency of analysis. It is lauched by running:

cd /guiding-optimisation-with-aiwc/codes beakerx --allow-root

from within the container and following the prompts to access it from the website front-end.

Note that if this node is accessed from an ssh session local ssh port forwarding is required and is achieved with the following:

ssh -N -f -L localhost:8888:localhost:8888 <node-name>