OpenMP help on code - c++

Question

this is my first time posting so I will appologise for my novice mistakes. Please also excuse the fact that not all variable names are in english. My problem is the following: i've written this code using openMP in both Visual Studio 2010 and in eclipse for c/c++ using the cygwin gcc compiler toolchain. In visual I get a speed-up but in eclipse I get a slow down twice the amount of the serial version. Can someone please explain what I have done wrong please? In short i'm just simulating the speed-up from when I copy from an array of 3D vectors into a double array in order to send over MPI.
#include <omp.h>
#include <time.h>
#include <stdio.h>
#include <vector>
const int NUMAR_FORME=10;
const int NUMAR_SECUNDE_SIMULATE=60; //number of buffers
const int dimensiuni_forme[10]={100,200,300,400,500,600,700,800,900,10000}; //size of each buffer
//-------- the buffers, cuurently only worker_buffer and buff is used
std::vector<std::vector<std::vector<double> > > worker_buffer;
std::vector<std::vector<double> > send_buffer,corect;
double **buff;
double **worker_buffer1;
long i,j,k,l;
int flag=0;
int numarator=0; //number of tests runed
clock_t start;
start=clock();
worker_buffer.resize(1);
buff = new double* [2];
int de_scris=0; //this tells me in which buffer to store, nou I alternate buff[0], buff[1], buff[0], buff[1]
worker_buffer[0].resize(NUMAR_SECUNDE_SIMULATE);
for(i=0;i<NUMAR_SECUNDE_SIMULATE;i++)
worker_buffer[0][i].resize(dimensiuni_forme[9]);
while(numarator<60)
{
if(numarator!=0)
delete [] buff[de_scris];
if(numarator!=0)
de_scris=(de_scris+1)%2;
long limita;
limita=NUMAR_SECUNDE_SIMULATE*dimensiuni_forme[9]*3; //3-comes from the fact that I will have a 3D vector structure
buff[de_scris]= new double [limita];
for(i=0;i<NUMAR_SECUNDE_SIMULATE;i++)
{ for(j=0;j<dimensiuni_forme[9];j++)
{
worker_buffer[0][i][j]=(i*dimensiuni_forme[9]+j)*3;
buff[de_scris][(i*dimensiuni_forme[9]+j)*3]=worker_buffer[0][i][j];
buff[de_scris][(i*dimensiuni_forme[9]+j)*3+1]=worker_buffer[0][i][j]+0.5;
buff[de_scris][(i*dimensiuni_forme[9]+j)*3+2]=worker_buffer[0][i][j]+0.75;
}
}
numarator++;
}
start=clock()-start;
printf("TICKS TOTAL %ld \n",start);
bool ad=true;
long nr;
for(i=0;i<NUMAR_SECUNDE_SIMULATE*dimensiuni_forme[9]*3;i++)
{
if(i%3==0)
nr=i;
if(i%3==0 && buff[de_scris][i]!=i)
ad=false;
else
if(i%3==1 &&buff[de_scris][i]!=(nr+0.5))
ad=false;
else
if(i%3==2 && buff[de_scris][i]!=(nr+0.75))
ad=false;
}
if(ad==false)
printf("not correct \n");
start=clock();
numarator=0;
//parallel version
while(numarator<60)
{
if(numarator!=0)
delete [] buff[de_scris];
long index, limita,id;
omp_set_num_threads(2);
if(numarator!=0)
de_scris=(de_scris+1)%2;
limita=NUMAR_SECUNDE_SIMULATE*dimensiuni_forme[9]*3; //3-
buff[de_scris]= new double [limita];
#pragma omp parallel shared(worker_buffer,limita,buff) private(index,id)
{
printf("intram cu %d threaduri \n", omp_get_num_threads());
id=omp_get_thread_num();
//index=id;
for(index=id;(index*3)<limita;index+=omp_get_num_threads())
{
buff[de_scris][index*3]=worker_buffer[0][index/dimensiuni_forme[9]][index%dimensiuni_forme[9]]; //aici va veni send_buff[index].x
buff[de_scris][index*3+1]=buff[de_scris][index*3]+0.5;
buff[de_scris][index*3+2]=buff[de_scris][index*3]+0.75;
}
// index+=omp_get_num_threads();
}//end parallel zone
numarator++;
}
start=clock()-start;
printf("TICKS TOTAL %ld \n",start);
ad=true;
//testing for correctness
for(i=0;i<NUMAR_SECUNDE_SIMULATE*dimensiuni_forme[9]*3;i++)
{
if(i%3==0)
nr=i;
if(i%3==0 && buff[de_scris][i]!=i)
ad=false;
else
if(i%3==1 &&buff[de_scris][i]!=(nr+0.5))
ad=false;
else
if(i%3==2 && buff[de_scris][i]!=(nr+0.75))
ad=false;
}
if(ad==false)
printf("not correct \n");
return 0;
}

Answer 1

Judging by how you organized this for loop:
for(index=id;(index*3)<limita;index+=omp_get_num_threads())
{
buff[de_scris][index*3]=worker_buffer[0][index/dimensiuni_forme[9]][index%dimensiuni_forme[9]]; //aici va veni send_buff[index].x
buff[de_scris][index*3+1]=buff[de_scris][index*3]+0.5;
buff[de_scris][index*3+2]=buff[de_scris][index*3]+0.75;
}
and assuming that you have 4 threads, your threads will get interleaved index values:
thread 0: 0, 4, 8, 12,...
thread 1: 1, 5, 9, 13,...
thread 2: 2, 6, 10, 14,...
thread 3: 3, 7, 11, 15,...
which may be causing cache ping-pong effects, since values written by different threads may land on the same cache line, thus slowing down your execution.
Try to use a simple for loop with static partitioning instead, in order to get continuous partitions:
#pragma omp parallel for
for(index = 0; index < limita / 3;index++)
{
buff[de_scris][index*3]=worker_buffer[0][index/dimensiuni_forme[9]][index%dimensiuni_forme[9]]; //aici va veni send_buff[index].x
buff[de_scris][index*3+1]=buff[de_scris][index*3]+0.5;
buff[de_scris][index*3+2]=buff[de_scris][index*3]+0.75;
}