Automatic parallelization in 4.3?

[Diego Novillo]

Janne Blomqvist wrote on 10/24/06 07:11:

> FWIW, I had a similar experience a few years ago when I tried the
> autoparallelization on XL Fortran/AIX. It was *a lot* slower than just
> using MPI over shared memory (ideological purity be damned).
> 
> Still waiting for the magic parallelization fairy, I suppose..
> 
This optimization will never be enabled by default.  Automatic 
parallelization belongs to a family of transformations that generally 
results in lowered performance.

The fundamental problem with autoparallelism is the need to preserve the 
sequential semantics of the code and the coarse granularity required to 
obtain performance.  Unlike ILP or data parallel transformations like 
vectorization, the glue code needed to spawn tasks and synchronization 
is significantly more expensive.

This does not mean that autoparallelism doesn't have its uses.  When it 
works, it can give you fairly decent speedups.  But the input code needs 
to have some special control and data flow properties that are not all 
that common in real world code.

To get the real benefit from multiprocessor machines, it is a lot more 
profitable to start thinking in terms of concurrent algorithms. 
Compiler switches on dusty-deck sequential code will only get you so far.