HR-Cache

This repo contains code for ICPE’24 paper: A Learning-Based Caching Mechanism for Edge Content Delivery

Installation and Running Guide

This guide provides instructions on how to build and run the project.

Prerequisites

Before you build and run this project, ensure you have the following prerequisites installed on your system:

• C++ Compiler: A compiler that supports C++17 standards (e.g., GCC or Clang).
• Git: Required for cloning repositories, including LightGBM.
• CMake: Needed for building LightGBM from source.
• pthread Library: For multi-threading support.
• cURL Library: Necessary for network operations in the server component.
• GNU Make: For processing the Makefile.
• LightGBM: Our project depends on LightGBM, a gradient boosting framework. Instructions for building LightGBM are provided below.

Building LightGBM

If LightGBM is not already installed and accessible by g++, follow these steps to build it from source:

mkdir -p builds
make build_lightgbm

This will build LightGBM from source.

Building the Project

After installing LightGBM, you can compile the project using the following command:

mkdir -p builds
mkdir executables
make hr

This will compile the project and create the necessary executables.

Running the Project

After building the project, you can run the simulation using the following command:

./executables/hr [args]

Replace [args] with the appropriate arguments for the simulation. The --file-path argument is required and specifies the path to the file to use for the simulation. To get paper results run as below with appropriate file and cache size:

./executables/hr --file-path=inputs/wiki2018.tr --cache-size=68719476736

Additional args:

• --file-path=: Specifies the path to the trace file to use for the simulation.
• --verbose: Enables verbose logging. If set to true, additional debug information will be printed.
• --rounds=: Specifies the number of rounds for the simulation.
• --cache-size=: Sets the cache size in bytes for the simulation.
• --evict-hot-for-cold: If set to true, cache-friendly object will be evicted for cache-averse object.
• --features-length=: Sets the length of the features for the simulation (2 is for frequency and size, to include 30 deltas for instance you would have to set this as 32).
• --hazard-bandwidth=: Specifies the hazard bandwidth for the simulation.
• --hazard-discrete: If set to true, the hazard will be discrete.
• --future-labeling: If set to true, look back labeling will be used.
• --one-time-training: If set to true, training will only be performed once.
• --max-boost-rounds=: Specifies the maximum number of boosting rounds.
• --feature-frequency: If set to true, frequency (not decayed frequency) will be used in features.
• --feature-decayed-frequency=: Sets the decay factor for feature decayed frequency.
• --feature-size: If set to true, size will be used as a feature.
• --report-interval=: Specifies the report interval for the simulation (by default this is 1 million).
• --log-file=: Specifies the name of the log file.

Trace Format:

Request traces are expected to be in a space-separated format with 3 columns:

• time (can be double)
• id should be a uint32, used to uniquely identify objects
• size should be uint32, this is object's size in bytes

time	id	size
1	1	140
3	2	290

Traces Used

We utilized four public traces to evaluate the performance of our method:

Wikipedia Datasets (wiki2018 and wiki2019)

• Source: The Wikipedia datasets for the years 2018 and 2019 were obtained from Learning Relaxed Belady for Content Distribution Network Caching, a paper by Zhenyu Song, Daniel S. Berger, Kai Li, and Wyatt Lloyd, presented at USENIX NSDI 2020.

EU Synthetic Dataset

• Source: The synthetic EU dataset was generated using the JEDI generator available at Jedi GitHub repository.

CloudPhysics Dataset

• Source: "Efficient MRC construction with SHARDS" by Carl A. Waldspurger, Nohhyun Park, Alexander Garthwaite, and Irfan Ahmad, presented at the 13th USENIX Conference on File and Storage Technologies (FAST '15), pages 95–110, Santa Clara, CA, February 2015.

The traces, as used in our analyses and experiments, are readily available in the input directory of this repository for ease of access and replication of our results.

Comparison with Other Algorithms

Our method has been compared against several algorithms to evaluate its performance effectively. These comparison algorithms are part of the LRB simulator.

For comprehensive analysis and understanding of how our method compares to these established algorithms, please see the simulation results and discussion sections of our paper.