[IPOL discuss] Nested openMP loops

[Miguel Colom]

Dear all,
We've detected some problems with demos that use nested loops with OpenMP.

Using a nested loop with openMP is always a bad idea since the outer  
loop will create as many threads as the inner loop iterations.

For example:
#pragma openmp parallel for
for = 1..10
   (...)
   #pragma openmp parallel for
   for j = 1..10
     ()...

This would create 10*10=100 threads if we configure OpenMP to allow  
nested loops. Of course, it doesn't make any sense since our server  
has just 32 cores.

Moreover, if the inner iterations execute fast, it might happen that  
the execution time of the iteration is less than the cost of  
assigning/releasing a thread.

In general, a good solution is to rewrite to code to use a single  
parallelized loop and precomputed indices. This way openMP can control  
that the number of threads it uses is related to the number of cores.  
And the performance is maximum.

Best,
Miguel