Huge slow down using openmp - parallel-processing

Question

I am trying to test the speed up for a small piece of code as follows:
for(i=0;i<imgDim;i++)
{
X[0][i] = Z[i] - U1[i] * rhoinv;
X[1][i] = Z[i] - U2[i] * rhoinv;
X[2][i] = Z[i] - U3[i] * rhoinv;
}
The iteration is around 200 and imgDim is 1000000. The total time for this piece of code is around 2 seconds. And the whole code cost about 15 seconds.
But after I use openmp to parallel this piece of code like:
omp_set_num_threads(max_threads);
#pragma omp parallel shared(X,Z,U1,U2,U3,imgDim,rhoinv) private(i)
{
#pragma omp for schedule(dynamic)
for(i=0;i<imgDim;i++)
{
X[0][i] = Z[i] - U1[i] * rhoinv;
X[1][i] = Z[i] - U2[i] * rhoinv;
X[2][i] = Z[i] - U3[i] * rhoinv;
}
}
max_threads is 8. Only this small piece of code needs around 11 seconds and the entire code use around 27 seconds. The most strange thing is the time decreases to 6 seconds if I change max_threads to 1. But still much longer than the sequential code.
It costs me a lot of time and I can not find the problem. Deeply appreciate if anyone can help me with that.

Answer 1

schedule(dynamic) introduces a huge run-time overhead. It should only be used for loops where each iteration could take a different amount of time and the improved load balancing would justify the overhead. For regular loops like yours dynamic scheduling is an overkill as it introduces unnecessary overhead, which slows down the computation.
Change the schedule type to static:
#pragma omp parallel for schedule(static)
for(i=0;i<imgDim;i++)
{
X[0][i] = Z[i] - U1[i] * rhoinv;
X[1][i] = Z[i] - U2[i] * rhoinv;
X[2][i] = Z[i] - U3[i] * rhoinv;
}
(Note: variables declared in outer scopes are shared by default and the parallel loop control variable is implicitly private)