[

]

Recently Michael Suess over at Thinking Parallel posted an interesting article on resolving Loop Carried Dependencies using OpenMP. These dependencies are so named because variables depend on previous iterations within a loop. If one wants to paralleli...

An important factor is how much time is spent in user code, and how much in system code. A time-like utility can be used to determine that.
The higher the proportion of time spent in user code, the higher the benefit of parallelization.

In this particular case, it seems memory access
opt [n] = Sn;
is the costly part, which reduces the benefit of parallelization.

1) there are some instances when parallelizing such loops makes sense - the classic case is to cause a side effect, like data distribution across multiple processors or to preserve a data distribution from previous loops. This particular case doesn't access much data, so the point is moot here.

2) It sometimes makes sense to parallelize loops with recurrences, even when the recurrence cannot be eliminated as in this case. Typical examples are when a LARGE otherwise fully parallel loop contains a small amount of recurrences, i.e., the ratio of parallel to synchronized code is large, especially on small numbers of processors.

Good post, Michael!

If you need more iterations for better statistics, why not reducing the multiplier 'up' from 1.1 to something like 1.001 (or even smaller)? If float overflows are your concern, this should allow for about 700000 iterations.

Bernd