Homomorphic Encryption Task Graph Mapping and Inter- and Intra-Chiplet Topology Generation Framework

This repository provides a Python-based framework for Design Space Exploration (DSE) of homomorphic encryption applications in multi-chiplet systems. It includes bilevel optimization algorithms, mapping and topology generation ablation experiments, and related experiment scripts.

Key Features

• Bilevel Optimization Algorithm: Optimize mapping and topology jointly.
• Baseline Ablations: Isolated experiments on mapping and topology for comparison.
• Flexible Input: Support for custom benchmark files, trace files, and topologies.
• Modular Structure: Easy to extend for new strategies or applications.

Hardware/Software Requirements

This framework was developed and tested in the following environment:

• Processor: 13th Gen Intel(R) Core(TM) i7-13650HX
• RAM: 32 GB DDR5
• Storage: 1TB NVMe SSD
• Operating System: Linux (Ubuntu 22.04.5 LTS)
• Software Dependency: Python 3.10.12

Getting Started

1. Clone Repository

git clone https://github.com/FCAS-LAB/FHE_ESWEEK.git
cd FHE_ESWEEK

2. Run CKKS_Resnet50 Demo (Bilevel Optimization Algorithm + Ablation Experiments)

bash run.sh

3. Run Individual Benchmarks

Using the commands below will execute all the experiments regarding CKKS_ResNet50, CKKS_CNN, CKKS_MLP, CKKS_VGG in this paper, including detailed bilevel optimization process and corresponding ablation experiments.

• CKKS_ResNet50:

python main.py \
  --time_file="./src/Resnet50/Resnet50_time.txt" \
  --bench_file="./src/Resnet50/Resnet50_bench.txt" \
  --trace_file="./src/Resnet50/Resnet50_trace.txt" \
  --topo_file="./src/Resnet50/Resnet50_gv.txt" \
  --show_file="./src/Resnet50/Resnet50_show.txt"

• CKKS_CNN:

python main.py \
  --time_file="./src/CNN/CNN_time.txt" \
  --bench_file="./src/CNN/CNN_bench.txt" \
  --trace_file="./src/CNN/CNN_trace.txt" \
  --topo_file="./src/CNN/CNN_gv.txt" \
  --show_file="./src/CNN/CNN_show.txt"

• CKKS_MLP:

python main.py \
  --time_file="./src/MLP/MLP_time.txt" \
  --bench_file="./src/MLP/MLP_bench.txt" \
  --trace_file="./src/MLP/MLP_trace.txt" \
  --topo_file="./src/MLP/MLP_gv.txt" \
  --show_file="./src/MLP/MLP_show.txt"

• CKKS_VGG:

python main.py \
  --time_file="./src/MLP/MLP_time.txt" \
  --bench_file="./src/MLP/MLP_bench.txt" \
  --trace_file="./src/MLP/MLP_trace.txt" \
  --topo_file="./src/MLP/MLP_gv.txt" \
  --show_file="./src/MLP/MLP_show.txt"

Output Interpretation

Bilevel Optimization Report

================ Running Bilevel Algorithm ================
# Iteration progress with time consumption updates
the original total time consumption: ... 
total time consumption after optimization iter 0: ...
total time consumption after optimization iter 1: ...
...
final optimized performance: ...  # Optimized execution time
final optimized topology (in the format of adjacency matrix):  # Optimized topology
1:[...]
2:[...]
3:[...]
...
90:[...]

Ablation Studies

================ Ablation Experiments ================
Mapping Ablation:
time consumption of random_mapping + ring-ring topology: ...
time consumption of proposed mapping + ring-ring topology: ...
time consumption of random mapping + mesh-mesh topology: ...
time consumption of proposed mapping + mesh-mesh topology: ...
-----------------------------------------------------
Topology Ablation:
...