DEAP Cache: Deep Eviction Admission and Prefetching for Cache

Ayush Mangal*, Jitesh Jain*, Keerat Kaur Guliani*, Omkar Bhalerao*

Contents

1. Overview
2. Setup Instructions
3. Repository Overview
4. Training and Testing
5. Results
6. License

1. Overview

This repo contains the code for the paper DEAP Cache: Deep Eviction Admission and Prefetching for Cache.

2. Setup Instructions

You can setup the repo by running the following commands:

$ git clone https://github.com/vlgiitr/deep_cache_replacement.git

$ pip install -r requirements.txt

3. Repository Overview

The repository contains the following modules:

• checkpoints/ - Contains the pretrained embeddings and a trained version of the DeepCache model.
• dataset/ - Dataset folder
  • address_pc_files/ - Contains csv files with addresses and PCs with their corresponding future frequency and recency
  • misses/ - Contains csv files with the missed (separately calculated for LRU and LFU) addresses and PCs with their corresponding future frequency and recency
• runs/ - Contains the tensorboard logs stored during DeepCache's training
• utils/ - Contains various utility files such as .py scripts for various baselines, etc.
• cache_lecar.py - Script for the modified LeCaR that evicts based on the future frequencies and recencies
• cache_model_train.py - Script for training the DeepCache model.
• create_train_dataset.py - Script for creating the dataloader for training DeepCache
• embed_lstm_32.py - Script for training the byte embeddings
• generate_binary_permutations.py - Script for generating a csv file with all the binary representations of numbers till 255 for the global vocabulary
• get_misses.py - Script for storing the missed addresses and PCs in csv files
• requirements.txt - Contains all the dependencies required for running the code
• standard_algo_benchmark.py - Script for caclculating hitrates on the dataset using all the baselines algorithms
• test_sim.py - Script for running the online test simulation

4. Training and Testing

• To train the byte-embeddings, run the following command:

$ python embed_lstm_32.py 

• To train DeepCache, run the following command:

$ python cache_model_train.py

• To run the online test simulation, run the following command

$ python test_sim.py

5. Results

The hit-rates for various baselines and our approach are given in the table below:

Method	Mean Hit-Rate
LRU	0.42
LFU	0.43
FIFO	0.36
LIFO	0.03
BELADY	0.54
Ours	0.48

It can be observed that our method comes the closest in performance to the optimal figure obtained from BELADY’s algorithm (Oracle), thus demonstrating the validity of our approach.

6. License

The code is released under MIT License.