You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

370 lines
15 KiB
C++

#include <mpi.h>
#include <fcntl.h>
#include <dirent.h>
#include <unistd.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <queue>
#include <ctime>
#include <string>
#include <vector>
#include <cstdio>
#include <fstream>
#include <cstdlib>
#include <iostream>
#include <algorithm>
/*
8-Byte numbers in 256KB = 32768
8-Byte numbers in 1MB = 131072
8-Byte numbers in 1GB = 134217728
All the programm assume numbers as 64_bits.
To visualize binary files in bash can be used:
od -t d8 -A n binaryfile.bin #For in use format
od -t d8 -A n --endian=little binaryfile.bin #For little-endian format
od -t d8 -A n --endian=big binaryfile.bin #For big-endian format
*/
9 months ago
#define BUFFERSIZE 64
#define RAMNUM 134217728
#define ALLOW_BUFFER 1
9 months ago
void sortedRuns(unsigned long long fileSize, unsigned long long sliceSize, unsigned long long maxLoop, FILE* file, int id, int mpiRank, int mpiSize)
{
unsigned long long startOffset, endOffset, currentOffset; //The interval is [startOffset, endOffset)
double startTot, start, end;
int64_t num;
std::vector<int64_t> bigVect;
11 months ago
int64_t buffer[static_cast<unsigned long long>(BUFFERSIZE)];
bigVect.reserve(sliceSize);
startTot = MPI_Wtime(); //Microsecond precision. Can't use time(), because each process will have a different "zero" time
start = MPI_Wtime();
for(unsigned long long l = 0; l < maxLoop; l++) //Populate the vector with the values in the file
{
startOffset = sliceSize * (mpiRank + (mpiSize * l));
if (startOffset >= fileSize)
break;
endOffset = startOffset + sliceSize;
fseek(file, startOffset * sizeof(int64_t), SEEK_SET);
currentOffset = startOffset;
bigVect.clear();
if (ALLOW_BUFFER) //Branch to test performance with and without buffer
{
while (currentOffset < endOffset)
{
11 months ago
unsigned long long elementsToRead = std::min(endOffset - currentOffset, static_cast<unsigned long long>(static_cast<unsigned long long>(BUFFERSIZE))); //It's important to check because if the difference between endOffset and startOffset is smaller than BUFFERSIZE we don't have to read further
9 months ago
unsigned long long elementsRead = fread(buffer, sizeof(int64_t), elementsToRead, file);//TODO
for (unsigned long long i = 0; i < elementsRead; ++i)
{
bigVect.push_back(buffer[i]);
}
currentOffset += elementsRead; //Increment currentOffset based on the number of elements read
11 months ago
if (elementsRead < static_cast<unsigned long long>(BUFFERSIZE)) // Check if we have reached the end of the file
break;
}
}
else
{
9 months ago
while (currentOffset < endOffset && fread(&num, sizeof(int64_t), 1, file) == 1)//TODO
{
bigVect.push_back(num);
currentOffset++;
}
}
end = MPI_Wtime();
9 months ago
std::cout << " " << end-start << "s" << " => Time to read file from offset " << startOffset << " to " << endOffset << " in Process " << mpiRank+1 << "/" << mpiSize << std::endl;
start = MPI_Wtime();
sort(bigVect.begin(), bigVect.end());
end = MPI_Wtime();
std::cout << " " << end-start << "s" << " => Time to sort elements in Process " << mpiRank+1 << "/" << mpiSize << " memory" << std::endl;
9 months ago
std::string templateName = "/mnt/raid/tmp/SortedRun" + std::to_string(id) + "_XXXXXX"; //If absolute path does not exist the temporary file will not be created (mandatory 6 times X)
int tmpFile = mkstemp(&templateName[0]); //Create a temporary file based on template
if (tmpFile == -1)
{
std::cout << "Error creating temporary file" << std::endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
9 months ago
start = MPI_Wtime();
for (unsigned long long i = 0; i < bigVect.size(); ++i) //Write the ordered number in a temp file
{
if (ALLOW_BUFFER) //Branch to test performance with and without buffer
{
11 months ago
buffer[i % static_cast<unsigned long long>(BUFFERSIZE)] = bigVect[i];
if ((i + 1) % static_cast<unsigned long long>(BUFFERSIZE) == 0 || i == bigVect.size() - 1)
{
9 months ago
ssize_t tw = write(tmpFile, buffer, sizeof(int64_t) * ((i % static_cast<unsigned long long>(BUFFERSIZE)) + 1));//TODO
if (tw == -1)
{
std::cout << "Error writing to file" << std::endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
}
}
else
{
int64_t elem = bigVect[i];
9 months ago
ssize_t tw = write(tmpFile, &elem, sizeof(int64_t));//TODO
if (tw == -1)
{
std::cout << "Error writing to file" << std::endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
}
}
9 months ago
end = MPI_Wtime();
off_t sz = lseek(tmpFile, 0, SEEK_END);
if (sz == 0)
{
if (close(tmpFile) == -1)
{
std::cout << "Error closing the file" << std::endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
if (unlink(&templateName[0]) == -1)
{
std::cout << "Error unlinking file" << std::endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
}
if (close(tmpFile) == -1)
{
std::cout << "Error closing the file" << std::endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
std::cout << " " << end-start << "s" << " => Time to write '" << templateName << "' and fill it up with " << sz/8 << " sorted elements by Process " << mpiRank+1 << "/" << mpiSize << std::endl;
start = MPI_Wtime();
}
end = MPI_Wtime();
9 months ago
std::cout << end-startTot << "s" << " => Time function sortedRuns() in Process " << mpiRank+1 << "/" << mpiSize << std::endl;
}
void kMerge(const std::string &argFile, int id, int mpiRank, int mpiSize)
{
std::string fileDir = "/mnt/raid/tmp/";
std::string pattern = "SortedRun" + std::to_string(id) + "_";
std::vector<int> fds; //To store the file descriptor of each file to merge
std::vector<std::string> fns; //To store the file name of each file to delete after merge
size_t lastSlash = argFile.find_last_of('/');
std::string nameOnly = (lastSlash != std::string::npos) ? argFile.substr(lastSlash + 1) : argFile;
std::string finalFile = "/mnt/raid/tmp/" + nameOnly + (ALLOW_BUFFER == 1 ? ".buf" : "") + ".sort";
double start, end;
int fileCount = 0;
DIR *dir = opendir(fileDir.c_str());
if (dir)
{
struct dirent *entry;
while ((entry = readdir(dir)) != nullptr)
{
if (entry->d_type == DT_REG) //Check if it's a regular file
{
std::string filename = entry->d_name;
if (filename.find(pattern) != std::string::npos) //Check if the file name matches the pattern
{
std::string tmpFile = fileDir + "/" + filename;
int fd = open(tmpFile.c_str(), O_RDONLY); //Open the file and save the file descriptor
if (fd != -1)
{
fds.push_back(fd);
fns.push_back(tmpFile);
fileCount++;
}
else
std::cout << "Error opening file '" << tmpFile << "' by Process " << mpiRank+1 << "/" << mpiSize << std::endl;
}
}
}
closedir(dir);
}
else
{
std::cout << "Error opening directory '" << fileDir << "' by Process " << mpiRank+1 << "/" << mpiSize << std::endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
9 months ago
int fdFinal = open(finalFile.c_str(), O_WRONLY | O_CREAT, S_IRUSR | S_IWUSR); //Open the file for writing only, creating it if it doesn't exist and (add '| O_EXCL' near O_CREAT if want the next feature) not overwrite if it exists
if (fdFinal == -1)
{
std::cout << "Error opening or creating final file '" << finalFile << "' by Process " << mpiRank+1 << "/" << mpiSize << std::endl;
9 months ago
for (const std::string &fn : fns) //Remove all temporary files
{
if (unlink(&fn[0]) == -1)
{
std::cout << "Error unlinking file '" << fn << "' by Process " << mpiRank+1 << "/" << mpiSize << std::endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
}
MPI_Abort(MPI_COMM_WORLD, 1);
}
std::cout << std::endl << "Starting the merge process for " << fileCount << " files" << std::endl << std::endl;
start = MPI_Wtime();
std::priority_queue<std::pair<int64_t, int>, std::vector<std::pair<int64_t, int>>, std::greater<std::pair<int64_t, int>>> minHeap; //Creating a Min Heap using a priority queue
int64_t tmpValue;
9 months ago
int b=0;
for(int i = 0; i < RAMNUM/fileCount; i++)
{
9 months ago
for (int fd : fds) //Populate the Min Heap with initial values from each file descriptor
{
9 months ago
switch (read(fd, &tmpValue, sizeof(int64_t)))
{
case sizeof(int64_t):
minHeap.push({tmpValue, fd});
break;
case 0:
b = 1;
break;
default:
std::cout << i << "Error reading size=" << sizeof(int64_t) << " from file descriptor by Process " << mpiRank+1 << "/" << mpiSize << std::endl;
MPI_Abort(MPI_COMM_WORLD, 1);
break;
}
}
9 months ago
if (b == 1)
break;
}
9 months ago
std::cout << minHeap.size() << "=size heap" << std::endl;
int tmpfd;
int64_t tmpValue2;
11 months ago
int64_t buffer[static_cast<unsigned long long>(BUFFERSIZE)];
unsigned long long i = 0;
while (!minHeap.empty()) //Write sorted elements to the temporary file
{
tmpValue = minHeap.top().first;
tmpfd = minHeap.top().second;
9 months ago
if (read(tmpfd, &tmpValue2, sizeof(int64_t)) == sizeof(int64_t)) //Read another integer from the same file descriptor//TODO
{
minHeap.pop();
minHeap.push({tmpValue2, tmpfd});
}
else //If no more values can be read
{
minHeap.pop();
9 months ago
if (fcntl(tmpfd, F_GETFD) == 1 && close(tmpfd) == -1)
{
std::cout << "Error closing the file descriptor by Process " << mpiRank+1 << "/" << mpiSize << std::endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
}
if (ALLOW_BUFFER) //Branch to test performance with and without buffer
{
11 months ago
buffer[i % static_cast<unsigned long long>(BUFFERSIZE)] = tmpValue;
if ((i + 1) % static_cast<unsigned long long>(BUFFERSIZE) == 0 || minHeap.empty())
{
9 months ago
ssize_t tw = write(fdFinal, buffer, sizeof(int64_t) * ((i % static_cast<unsigned long long>(BUFFERSIZE)) + 1));//TODO
if (tw == -1)
{
std::cout << "Error writing to file" << std::endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
}
i++;
}
else
{
9 months ago
ssize_t tw = write(fdFinal, &tmpValue, sizeof(int64_t));//TODO
if (tw == -1)
{
std::cout << "Error writing to file" << std::endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
9 months ago
}
}
for (const std::string &fn : fns) //Remove all temporary files after merging them
{
if (unlink(&fn[0]) == -1)
{
std::cout << "Error unlinking file '" << fn << "' by Process " << mpiRank+1 << "/" << mpiSize << std::endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
}
end = MPI_Wtime();
std::cout << end-start << "s" << " => Time function kMerge() in Process " << mpiRank+1 << "/" << mpiSize << std::endl;
std::cout << std::endl << "Sorted file '" << finalFile << "'" << std::endl;
}
int main(int argc, char* argv[])
{
MPI_Init(&argc, &argv); //Initialize the MPI environment
double startGlobal, endGlobal;
int id, mpiSize, mpiRank;
MPI_Comm_size(MPI_COMM_WORLD, &mpiSize); //Get the number of processes
MPI_Comm_rank(MPI_COMM_WORLD, &mpiRank); //Get the number of process
if (mpiRank == 0)
{
startGlobal = MPI_Wtime();
std::srand(std::time(0));
id = std::rand() % 10000; //Get a random id number to recognize files of different executions
}
MPI_Bcast(&id, 1, MPI_INT, 0, MPI_COMM_WORLD);
if (argc != 2)
{
if (mpiRank == 0)
{
std::cout << "Usage: " << argv[0] << " <file_to_parse>" << std::endl;
9 months ago
std::cout << "It returns a file with extension '.sort' in the /mnt/raid/tmp/ directory. Make sure to have space before." << std::endl;
std::cout << "Use arguments in the make as ARGS=\"stuff\". Example 'make run ARGS=\"/path/to/file\"'." << std::endl;
}
MPI_Finalize(); //Clean up the MPI environment
return 0;
}
FILE *file;
unsigned long long fileSize, slices, sliceSize, maxLoop;
file = fopen(argv[1], "rb"); //Open the file in mode rb (read binary)
if(!file)
{
std::cout << "Error opening file: " << argv[1] << std::endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
fseek(file,0,SEEK_END);
fileSize = ftell(file) / 8; //Size in bytes of the file, correspond to the number of numbers to parse. Each number is 8 bytes
if (mpiRank == 0)
9 months ago
std::cout << "Using " << mpiSize << " nodes for sorting " << (fileSize/134217728 > 1 ? fileSize/134217728 : fileSize/131072) << (fileSize/134217728 > 1 ? "Gb" : "Mb") << " file '" << argv[1] << "' of " << fileSize << " elements" << std::endl << std::endl;
11 months ago
//Load balancer
9 months ago
if (fileSize < ((unsigned long long) static_cast<unsigned long long>(RAMNUM) * mpiSize)) //Can add more granularity considering efficiency, now is used by default all nodes
11 months ago
slices = mpiSize;
else
11 months ago
slices = mpiSize;
9 months ago
sliceSize = (fileSize / slices); //Each process divides a number of 8-byte integers based on the size of the starting file
11 months ago
maxLoop = 1;
if (sliceSize > static_cast<unsigned long long>(RAMNUM))
9 months ago
{
maxLoop = (fileSize / (static_cast<unsigned long long>(RAMNUM) * mpiSize)) + 1;
sliceSize = (fileSize / maxLoop);
sortedRuns(fileSize, sliceSize + 1, maxLoop, file, id, mpiRank, mpiSize);
}
else
9 months ago
{
sortedRuns(fileSize, sliceSize + 1, maxLoop, file, id, mpiRank, mpiSize);
}
fclose(file);
MPI_Barrier(MPI_COMM_WORLD); //Blocks the caller until all processes in the communicator have called it
if(mpiRank==0)
{
kMerge(argv[1], id, mpiRank, mpiSize);
endGlobal = MPI_Wtime();
std::cout << (endGlobal-startGlobal)/60.0 << "min" << " => FULL EXECUTION TIME" << std::endl;
}
MPI_Finalize(); //Clean up the MPI environment
return 0;
}
11 months ago