From serial to omp: no speedup - c++

Question

I'm new to openMP. I'm working on All Pair Shortest Path Algorithm and here is the serial C++ routine i need to parallelize (complete code at the end of the post):
void mini(vector<vector<double>> &M, size_t n, vector<double> &rowk, vector<double> &colk)
{
size_t i, j;
for ( i=0; i<n; i++)
for ( j=0; j<n; j++)
M[i][j]=min(rowk[j]+colk[i], M[i][j]);
}
At execution I get this :
$ time ./floyd
real 0m0,349s
user 0m0,349s
sys 0m0,000s
Now, I try to insert some directives:
void mini(vector<vector<double>> &M, size_t n, vector<double> &rowk, vector<double> &colk)
{
#pragma omp parallel
{
size_t i, j;
#pragma omp parallel for
for ( i=0; i<n; i++)
for ( j=0; j<n; j++)
M[i][j]=min(rowk[j]+colk[i], M[i][j]);
}
}
Unfortunately, there is no speedup:
$ grep -c ^processor /proc/cpuinfo
4
$ time ./floyd
real 0m0,547s
user 0m2,073s
sys 0m0,004s
What am I doing wrong?
EDIT
Processor: Intel(R) Core(TM) i5-4590 CPU (4 hardware cores)
Complete code :
#include <cstdio>
#include <vector>
#include <limits>
#include <ctime>
#include <random>
#include <set>
#include <omp.h>
using namespace std;
typedef struct Wedge
{
int a, b;
double w;
} Wedge;
typedef pair<int, int> edge;
int randrange(int end, int start=0)
{
random_device rd;
mt19937 gen(rd());
uniform_int_distribution<> dis(start, end-1);
return dis(gen);
}
void relax_omp(vector<vector<double>> &M, size_t n, vector<double> &rowk, vector<double> &colk)
{
#pragma omp parallel
{
size_t i, j;
#pragma omp parallel for
for (i=0; i<n; i++)
for ( j=0; j<n; j++)
M[i][j]=min(rowk[j]+colk[i], M[i][j]);
}
}
void relax_serial(vector<vector<double>> &M, size_t n, vector<double> &rowk, vector<double> &colk)
{
size_t i, j;
for (i=0; i<n; i++)
for ( j=0; j<n; j++)
M[i][j]=min(rowk[j]+colk[i], M[i][j]);
}
void floyd(vector<vector<double>> &dist, bool serial)
{
size_t i, k;
size_t n {dist.size()};
for (k=0; k<n; k++)
{
vector<double> rowk =dist[k];
vector<double> colk(n);
for (i=0; i<n; i++)
colk[i]=dist[i][k];
if (serial)
relax_serial(dist, n, rowk, colk);
else
relax_omp(dist, n, rowk, colk);
}
for (i=0; i<n; i++)
dist[i][i]=0;
}
vector<Wedge> random_edges(int n, double density, double max_weight)
{
int M{n*(n-1)/2};
double m{density*M};
set<edge> edges;
vector<Wedge> wedges;
while (edges.size()<m)
{
pair<int,int> L;
L.first=randrange(n);
L.second=randrange(n);
if (L.first!=L.second && edges.find(L) == edges.end())
{
double w=randrange(max_weight);
Wedge wedge{L.first, L.second, w};
wedges.push_back(wedge);
edges.insert(L);
}
}
return wedges;
}
vector<vector<double>> fill_distances(vector<Wedge> wedges, int n)
{
double INF = std::numeric_limits<double>::infinity();
size_t i, m=wedges.size();
vector<vector<double>> dist(n, vector<double>(n, INF));
int a, b;
double w;
for (i=0; i<m; i++)
{ a=wedges[i].a;
b=wedges[i].b;
w=wedges[i].w;
dist[a][b]=w;
}
return dist;
}
int main (void)
{
double density{0.33};
double max_weight{200};
int n{800};
bool serial;
int ntest=10;
double avge_serial=0, avge_omp=0;
for (int i=0; i<ntest; i++)
{
vector<Wedge> wedges=random_edges(n, density, max_weight);
vector<vector<double>> dist=fill_distances(wedges, n);
double dtime;
dtime = omp_get_wtime();
serial=true;
floyd(dist, serial);
dtime = omp_get_wtime() - dtime;
avge_serial+=dtime;
dtime = omp_get_wtime();
serial=false;
floyd(dist, serial);
dtime = omp_get_wtime() - dtime;
avge_omp+=dtime;
}
printf("%d tests, n=%d\n", ntest, n);
printf("Average serial : %.2lf\n", avge_serial/ntest);
printf("Average openMP : %.2lf\n", avge_omp/ntest);
return 0;
}
output :
20 tests, n=800
Average serial : 0.31
Average openMP : 0.61
command line:
g++ -std=c++11 -Wall -O2 -Wno-unused-result -Wno-unused-variable -Wno-unused-but-set-variable -Wno-unused-parameter floyd.cpp -o floyd -lm -fopenmp

Answer 1

Your main issue is that you accidentally use nested parallelism:
#pragma omp parallel
{
size_t i, j;
#pragma omp parallel for
Since you already are in a parallel region, your second line should be
#pragma omp for
Otherwise, since a omp parallel for equals a omp parallel and a omp for, you have two nested parallel regions which is typically bad. Fixing this minor thing gets an ~2x speedup on a similar CPU.
There are several limitations why you are unlikely to get a full 4x speedup, such as but not limited to:
Memory bandwidth as a bottleneck
Relative overhead due to the small amount of work done within the parallel loop
Lower clock frequencies with multiple threads in turbo mode
Edit:
By the way, the much more idiomatic way to write your code is the following:
void relax_omp(...) {
#pragma omp parallel for
for (size_t i=0; i<n; i++) {
for (size_t j=0; j<n; j++) {
M[i][j]=min(rowk[j]+colk[i], M[i][j]);
}
}
}
If you declare variables as locally as possible, OpenMP wil almost alaways do the right thing. Which, in this case, means that i and j are private. In general it is much easier to reason about code this way.

Answer 2

There could be many reasons for this, the most obvious being that the work load is too small to notice speed up. The initial work load is 300ms. I would suggest enclosing this in a serial outerloop that repeats this work for at least 20 times, then you are starting with a serial time of (300ms * 20) 6 seconds to test with.
The other factor is the availability of parallel cores on the machine you are running this on. If your cpu has one core, multi-threading will cause slowdown due to the cost of thread-switching. 2 logical cores should show some speed up, 2 physical cores may show close to linear speed up.
Using pragma directives alone also does not guarantee that openMP is used. You have to compile using the -fopenmp command line argument to guarantee that the openmp library is linked to your object code.
Edit
Looking at your code now, the factor that controls the amount of work seems to be N rather than the outer loop. The idea of the outer loop was to artificially increase the amount of work within the same timing period but that can't be done here as you are trying to solve a specific problem. You can try parallelizing the nested loop as well but I think N = 800 is too low for parallelization to make a difference.
#pragma omp parallel for private(j) collapse(2)
j needs to be private to each iteration of the outer loop, hence private(j), otherwise j gets shared across all threads, leading to an inaccurate result.
Your loop is executed 640,000 times which is not much for modern CPUs that clock 3GHZ+, try something around N = 5000 which is 25M iterations.