Read Binary Files

In this tutorial file we show to how to read a binary file in parallel with PIGO.

The Setup

Suppose we are reading a dense array of integers from a binary file. The file format is simple: first the number of integers are stored (as an unsigned integer) and then, one after another, the integers.

We have an existing program that reads such arrays and the goal is to modify it to read with PIGO to take advantage of parallel improvements.

The full format is as follows:

<size:uint64_t><data:int32>...

On Linux with bash, we can generate a random large file with the following commands (Warning: this will create a 3.8 GB file. Adjust the numbers if necessary.):

perl -e 'print(pack('Q',999817214))' > ints.bin
dd if=/dev/urandom of=ints.bin seek=1 bs=8 count=131071
dd if=/dev/urandom of=ints.bin seek=1 bs=1048576 count=3813

Loading the Vector

We want to build a program that loads the vector to then perform some action on it.

Following is the existing program that we start with (readbin.cpp)

#include <iostream>
#include <fstream>
#include <cstdint>
#include <omp.h>
using namespace std;

void do_something(uint64_t size, int32_t* vec) { }

int main(int argc, char** argv) {
    // Open and load the file
    if (argc != 2) {
        cerr << "Usage: " << argv[0] << " file" << endl;
        return 1;
    }

    double t = omp_get_wtime();
    uint64_t size;

    // Open the input file
    auto fin = fstream(argv[1], ios::in | ios::binary);
    // Read the expected size of the file
    fin.read((char*)&size, sizeof(size));
    // Allocate the vector
    int32_t* vec = new int32_t[size];
    // Copy the integers
    fin.read((char*)vec, sizeof(int32_t)*size);

    cerr << "[read] " << omp_get_wtime()-t << endl;

    do_something(size, vec);

    // Cleanup
    delete [] vec;
    fin.close();
    return 0;
}

This program can be compiled on Linux with g++ -std=c++11 -O3 -o readbin readbin.cpp -fopenmp and on Mac, after brew install libomp, with clang++ -std=c++11 -O3 -o readbin readbin.cpp -Xclang -fopenmp /usr/local/lib/libomp.dylib.

Loading Faster

We can use PIGO to load in parallel instead. Ensure that pigo.hpp is in the same folder.

PIGO has support for reading and writing binary files through the FileReader. This class ROFile can be used for read-only files. First, create an ROFile with the filename to open using ROFile fin { argv[1] };. Next, the uint64_t size can be read out by using the read method of the ROFile, size = fin.read<uint64_t>(); Finally, the vector can be populated by replacing the read with a parallel_read, which takes the destination and the number of bytes to read.

The modified program is as follows:

#include <iostream>
#include <fstream>
#include <cstdint>
#include <omp.h>
using namespace std;

#include "pigo.hpp"
using namespace pigo;

void do_something(uint64_t size, int32_t* vec) { }

int main(int argc, char** argv) {
    // Open and load the file
    if (argc != 2) {
        cerr << "Usage: " << argv[0] << " file" << endl;
        return 1;
    }

    // Read the input file
    double t = omp_get_wtime();
    ROFile fin { argv[1] };
    uint64_t size = fin.read<uint64_t>();
    int32_t* vec = new int32_t[size];
    fin.parallel_read((char*)vec, sizeof(int32_t)*size);
    cerr << "[read] " << omp_get_wtime()-t << endl;

    do_something(size, vec);

    delete [] vec;
    return 0;
}

Performance Comparison

Following are the results in seconds between the original, fread version and the PIGO version. (C) means the cache is cold and (W) means warm.

System	fread	PIGO	Speedup
Laptop (C)	6.44	6.39	1.01 x
Laptop (W)	3.25	1.48	2.20 x
Workstation (C)	4.22	1.50	2.81 x
Workstation (W)	2.32	0.26	8.92 x
Server (C)	6.53	1.53	4.27 x
Server (W)	1.73	0.18	9.61 x