GreenRec: Carbon-aware Recommendation System with Approximate Computation Reuse

Contents

• 1. Code Structure
• 2. About simulation
• 3. Experimental Setup
• 4. Getting Started Instructions
• 5. Run Simulator
• 6. Results
• 7. Appendix

1. Code Structure

Folders	Description
bashrc	Contains simulation running bash script code (simulator.sh)
data	Preprocessed carbon intensity and recommendation workload trace file
result	Simulation logs and results are stored in this folder
saved	DNN recommendation models should be placed in this folder
src	All GreenRec function and simulator Python codes

2. About simulation

This repository provides the code and data to compare the performance of GreenRec with BASE. By simply running a bash script, you can simulate various environments (such as carbon intensity, recommendation workload traces, and accuracy constraints). Most of the implementation follows what is introduced in the paper. However, for the energy measurement module, profiled values are used for compatibility. If you want to perform actual energy measurements, you can use the functions in the src/ENERGY_function.py file.

3. Experimental Setup

The experimental environment is configured as follows:

Component	Specification
Processor	Intel Xeon Gold 5218 2.3 GHz, 64-Core
Chipset	Intel C621
Memory	DDR4 2666 MHz, 128 GB (32 GB x4)
GPU	NVIDIA RTX 2090
Driver	NVIDIA Driver 545.23.08
CUDA	cuda_12.3.r12.3
OS	Ubuntu 20.04 Server (kernel v5.16.1)

4. Getting Started Instructions

For compatibility, there is a Dockerfile that can imitate our evaluation setup.

First, build Docker image.

docker build -t dockerID/greenrec:tag

Second, run Docker file and place GreenRec repository folder inside the docker container.

docker run -dit --privileged --name GreenRec --gpus all dockerID/greenrec:tag
docker cp GreenRec:path GreenRec_repository_folder

Third, install all requirements to run GreenRec.

pip install -r requirements.txt

5. Run Simulator

Running the bashrc/simulator.sh script starts the simulation. Caution: you should execute simulator.sh inside the bashrc folder since our repository uses relative paths in simulator code.

bash ./simulator.sh arguments

The simulator.sh script requires five arguments, each of which has the following meaning:

Argument	Description
1. Simulation Mode	0: BASE 1: GreenRec
2. Recommendation Workload	0: MovieLens 1: Twitch
3. Carbon Intensity	0: NESO Increase 1: NESO Decrease 2: NESO Stable
4. Accuracy Constraint	0: Const. High 1: Const. Mid 2: Const. Low
5. Refine parameter	0 ~ 320

For example, if you want to simulate MovieLens workload with GreenRec on NESO Decrease, Mid Constraint and refine parameter 40, you should complete your running code like this.

bash ./simulator.sh 1 0 1 1 40

By this way, simulator code starts running and writes logs and results in result folder

6. Results

If all simulations terminates, in log file, you can see final result.

================== simulation result ====================
TOTAL CARBON: 44.195512771536244
OPTIMIZE CARBON: 4.267921061571182
SEARCH CARBON: 0.9419837009668395
EXTRACT CARBON: 2.168440940745376
INSERT CARBON: 0.3015082769288639
SAVE CARBON: 0.0002531014664074795
INFERENCE CARBON: 36.51540568985581


CACHE HIT RATE: 0.44843322178278067
HIT10 : 0.5372680255552176
Accuracy Constraint Guarantee: 1.0129922270104363

According to the result, GreenRec can operate with low carbon footprint while maintaining considerable accuracy.

7. Appendix

Raw Datasets and vanilla DNN recommendation model code used in this simulation is accessible in these links

asset	Link
MovieLens-1M	https://grouplens.org/datasets/movielens/
Twitch-100K	https://github.com/JRappaz/liverec?tab=readme-ov-file
Bert4Rec	https://github.com/FeiSun/BERT4Rec