We support two types of evaluation for PrivateKube and DPF (see OSDI paper for methodologies):
microbenchmark
: uses a simulator of the privacy scheduling algorithm (such as DPF) under a highly controlled workload;macrobenchmark
: uses the real system under a DP ML workload we developed over the Amazon Reviews dataset.
This folder provides code and instructions for how to perform/reproduce both types of evaluation. Instructions of how to reproduce the macrobenchmark results are in ../README.md. Instructions for how to reproduce the microbenchmark results are in ./microbenchmark/README.md.