Rysia is a declartive benchmarking framework for deep learning systems. Users specify deep learning workloads in blueprint files without any coding required. Compilation into platform specific execution is done automatically. Benchmarking experiments can be executed locally or in an AWS cloud environment.

To install CPU-restricted packages, run

python setup_cpu.py install

To install GPU-accelerated packages, run

python setup_gpu.py install

Experiment worklaods are specified fully declarative in blueprint files which are implemented as Python modules. Examples can be found in the blueprints folder. An overview over available parameters can be found in the blueprint_example.py file.

To execute benchmarking workloads locally:

1. Install the package
2. Specify a blueprint file and set the parameter aws to False
3. Run rysia blueprint.py

To run benchmarking workloads within the AWS cloud environment:

1. Install the awscli and configure a valid AWS account
2. Build and push the package to an Amazon ECR repository by specifying the required parameters in build_and_push.py and then running python build_and_push.py
3. Specify job definitions that are linked to the ECR container images in the AWS Batch web terminal
4. Specify a blueprint file and set the parameter aws to True
5. Run rysia blueprint.py