This repo is for the SC 2022 artifacts evaluation.

• If you are using the server that we provided, you don't need to setup the environment and prepross the dataset, please use the docker container SC_artifacts and SC_ae_hugectr (only for running hugectr) to run the scripts.
• If you want to run the codes on your own server, please refer to the [Docker Image](# Docker-Image) , [Data Process](# Data-Process) and [Generate Index Bijection](# Generate-Index-Bijection) sections.

We provoded some script to run the expriment and get the result in our paper, includes:

• Figure 11:
  • We provide codes for running DLRM (CPU-GPU), TT-Rec and EL-Rec.
  • Please also refer to the Figure11/README.md.
• Figure 12:
  • We provide codes for running DLRM (1 GPU), DLRM (4 GPU), EL-Rec (1 GPU) and EL-Rec (4 GPU).
  • Please also refer to the Figure12/README.md.
• Figure 13:
  • We provide codes for running EL-Rec (1/2/4 GPU), TorchRec(2/4 GPU) and HugeCTR (2/4 GPU).
  • To run the EL-Rec, please use the docker container SC_artifacts.
  • To run the HugeCTR, please use the docker container SC_ae_hugectr.
  • To run the HugeCTR, please use the conda environment new_torchrec.
  • Please also refer to the Figure13/README.md.
• Figure 14:
  • We provide codes to get the breakdown study of EL-Rec.
  • Please also refer to the Figure14/README.md.
• Figure 15:
  • We provide codes to draw the loss convergence curve of DLRM, TT-Rec and EL-Rec.
  • Please also refer to the Figure15/README.md.
• Figure 16:
  • We provide codes for running DLRM, EL-Rec (Sequential) and EL-Rec (Pipeline).
  • Please also refer to the Figure16/README.md.
• Figure 17:
  • We provide codes to get the breakdown study of Efficient TT-table lookup.
  • Please also refer to the Figure17/README.md.
• Figure 18:
  • We provide codes to get the breakdown study of Efficient TT-table backward.
  • Please also refer to the Figure18/README.md.
• Table 4:
  • We provide codes to get the test accuracy of different frameworks.
  • Please also refer to the Table4/README.md.

We need two different dockers to reproduce the evaluation results.

docker pull happy233/zheng_dlrm:latest
docker pull zhengwang0122/dlrm_hugectr:latest

To run the docker images, please using the following commands:

docker run --gpus=all -it --cap-add SYS_NICE -v $$<your folder>$$:/workspace/SC_artifacts_eval -w /workspace/SC_artifacts_eval happy233/zheng_dlrm:latest
docker run --gpus=all -it --cap-add SYS_NICE -v $$<your folder>$$/Figure13:/workspace/HugeCTR -w /workspace/HugeCTR zhengwang0122/dlrm_hugectr:latest

Examples:

docker run --gpus=all -it --cap-add SYS_NICE -v /home/zhengw/workspace/SC_artifacts_eval:/workspace/SC_artifacts_eval -w /workspace/SC_artifacts_eval happy233/zheng_dlrm:latest
docker run --gpus=all -it --cap-add SYS_NICE -v /home/zhengw/workspace/SC_artifacts_eval/Figure13:/workspace/HugeCTR -w /workspace/HugeCTR zhengwang0122/dlrm_hugectr:latest

Please install the following packages in happy233/zheng_dlrm image:

pip install sympy

Download dataset:

• Kaggle: https://figshare.com/articles/dataset/Kaggle_Display_Advertising_Challenge_dataset/5732310
• Terabyte: https://labs.criteo.com/2013/12/download-terabyte-click-logs/
• Avazu: https://www.kaggle.com/c/avazu-ctr-prediction/data

Unzip and save dataset to:

dlrm_dataset/
├── avazu
│   └── avazu
├── kaggle
│   ├── readme.txt
│   ├── test.txt
│   └── train.txt
└── terabyte
    ├── day_0
    ├── day_1
    └── day_2

First, make directory:

mkdir processed_data/avazu
mkdir processed_data/avazu/processed
mkdir processed_data/kaggle
mkdir processed_data/kaggle/processed
mkdir processed_data/terabyte/processed
mkdir processed_data/workspace

Run script to process data:

cd data_process
./run.sh

First install rabbit_module:

cd /workspace/SC_artifacts_eval/rabbit_module/src
apt-get install libboost-all-dev
apt-get install libgoogle-perftools-dev
python setup.py install

Then run script to record access pattern:

cd /workspace/SC_artifacts_eval/rabbit_module/
./1_record_dataset.sh

Last step, generate index bijection and save the generate dataset for trianing:

cd /workspace/SC_artifacts_eval/rabbit_module/
./2_index_bijection_generate.sh
./3_save_training_data.sh

We use some codes from the following repos:

• Facebook DLRM
• FAE
• TT-Rec
• TorchRec
• HugeCTR