Introduction

This repository contains a comprehensive set of IO benchmarks for Linux and OS X. These benchmarks are intended to determine the fastest ways to perform the following IO operations on each target platform:

Sequentially reading a file.
Sequentially overwriting a preallocated file.
Replacing the contents of an existing file with those of another file.

Prerequisites

Linux (version 2.6.33 or newer) or OS X.
A C++11-conformant compiler that accepts flags in GCC's format (e.g. g++ or clang++).
Boost.
Ruby.
Rake.

Compilation

Before compiling the benchmarks, you will need to set two environment variables.

Set CXX to your desired C++11-conformant compiler.
Set BOOST_INCLUDE_PATH to the path containing the Boost header files.

Afterwards, you can compile the benchmarks by running rake.

Usage

The tools directory contains a set of scripts that you will need to run the benchmarks. In order to run these scripts, you will need to type chmod +x tools/*.

The tools/make_data.rb script uses dd to create a set of files in the data directory. These files are used to perform the benchmarks.
The tools/test_read.sh and tools/test_write.sh scripts perform the reading and writing benchmarks, respectively.
The read benchmark must be run as root! This is because the benchmark repeatedly drops the page cache to obtain accurate results. Do not run this benchmark on a server that is doing anything important!
I did not create a script to run the copy benchmark. Based on existing results, it is clear that the fastest way to copy a file on OS X is copy_mmap, and splice_preallocate_fadvise or sendfile_preallocate_fadvise on Linux.

In both test_read.sh and test_write.sh, you will see the following lines:

#sizes=(8 16 24 32 40 48 56 64 80 96 112 128 160 192 224 256 320 384 448 512 640 768 896 1024)
sizes=(8 16 32 64 80 96 112 256) #512 1024)

These lines declare the sizes of the files (in megabytes) that are used for the benchmarks. By default, only a small set of files ranging in size from 8 MB to 256 MB are used. The first line in the pair refers to the full set of files produced by tools/make_data.rb. Even on a machine with a fast PCIe SSD, the read benchmark did not finish overnight. So only uncomment this line if you know that you will be able to leave the benchmark running for a long time (half a day to several days, depending on the speed of your hard drive).

The results of the benchmarks are saved in the results directory. This directory already contains results generated from a couple of systems.

License

This project is released under the Creative Commons Attribution 4.0 International License.

References

The manual pages for OS X and Linux.
A very useful benchmark on write patterns.
The blog post by the same author.
A Mozilla blog post about F_PREALLOCATE on OS X.

adityaramesh / io_benchmark