Re: [Gomp-discuss] Release candidate

[Scott Robert Ladd]

Lars Segerlund wrote:
> Get it in there and get some feedback :-) that's my opinion.
>
> OpenMP is alredy mentioned on the gfortran pages, so I think were
> already made a bit of a stir.

Things are stirring elsewhere; give me a spoon, and I'll use it!

> I disagree slightly about the libs and their design, ( based on
> looking at intels openMP compiler which handles threads a bit
> different in order to get higher performance on the numa parts, and
> my wishes to be able to run on MPI or something similar, there are a
> lot of clusters showing up nowadays and I think we need so supprot
> them ). ( Ok another shameless plug for my humble opinions :-) ).

We already have MPI and PVM as favored cluster programming tools; many 
people are interested in hybrids, where OpenMP is used in conjunction 
with MPI. I realize the OdinMP allowed MPI, but I don't think that is 
the intent of the OpenMP design, nor the best use of limited resources.

> I'm doing some stuff in gfortran which indirectly relates to openMP
> in that it's parser and backend related, and have been doing some
> small tests on the fortran frontend and it's handling of comments in
> order to see if the interface I have grafted onto gfortran also can
> be used for the openmp directives, so far it looks ok, I have a
> couple of show stoppers :-) .. ( nothing bad but ! ).

Work needs to be coordinated between the different frontends. As I said 
in my earlier message, I'll have a complete requirements document up in 
the next day or so for comment, with a design document in development by 
mid-November.

> The approach I have is to hack a generic plugin for directives in
> comments, and then all our code would be selfcontained, this is
> fairly easy to do with the fortran parser so I hope it will pay off. 

I'd prefer to design, as opposed to hack. I realize we're all anxious.

> I think we shall keep as many options open as possible while
> remembering the words of Linux Thorvalds, never overdesign, get
> something running first, it's more important. If we need changes
> later we can just rip it apart and stick it together again.

His name or "Torvalds." ;) And I'm quite the admirer of Linus -- but he 
has a true dictatorship in place, at least in terms of the vanilla 
kernel. People can hack all they want, but the buck stops with Linus. 
Kernel developers spend an awful lot of time arguing, backtracking, and 
undoing that might be avoided by a bit more design and a bit less hacking.

It's a matter of balance; too much design, and all you have is 
documents. Too much hacking, and you spend time tearing down and rebuilding.

--
Scott Robert Ladd
site: http://www.coyotegulch.com
blog: http://chaoticcoyote.blogspot.com